Machine and deep learning applied to medical microwave imaging: a scoping review from reconstruction to classification

Prog. Biomed. Eng. 7(2025) 042008 https://doi.org/10.1088/2516-1091/ae0bd3 Progress in Biomedical Engineering OPEN ACCESS RECEIVED 14 April 2025 REVISED 22 August 2025 ACCEPTED FOR PUBLICATION 25 September 2025 PUBLISHED 15 October 2025 Original content from this work may be used under the terms of the Creative Commons Attribution 4.0 licence. Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. TOPICAL REVIEW Machine and deep learning applied to medical microwave imaging: a scoping review from reconstruction to classification Tiago M M Silva1, Raquel C Conceiç˜ ao1,2and Daniela M Godinho1,2,∗ 1Departamento de Física, Faculdade de Ciˆ encias, Universidade de Lisboa, 1749-016 Lisbon, Portugal 2Instituto de Biofísica e Engenharia Biomédica, Faculdade de Ciˆ encias, Universidade de Lisboa, 1749-016 Lisbon, Portugal ∗Author to whom any correspondence should be addressed. E-mail: [email protected] Keywords: machine learning, deep learning, classification, image reconstruction, microwave imaging, breast cancer, brain diseases Abstract Microwave imaging (MWI) is a promising modality due to its non-invasive nature and lower cost compared to other medical imaging techniques. These characteristics make it a potential alternative to traditional imaging techniques. It has various medical applications, particularly explored in breast and brain imaging. Machine learning (ML) has also been increasingly used for medical applications. This paper provides a scoping review of the role of ML in MWI, focusing on two key areas: image reconstruction and classification. The reconstruction section discusses various ML algorithms used to enhance image quality, highlighting methods such as convolutional neural network and support vector machine. The classification section delves into the application of ML for distinguishing between different tissue types, including applications in breast cancer detection and neurological disorder classification. By analyzing the latest studies and methodologies, this review addresses the current state of ML-enhanced MWI and sheds light on its potential for clinical applications. 1. Introduction Microwave imaging (MWI) is defined as the image of the internal composition of an object using electromagnetic fields within the microwave frequency range from 300 MHz to 30 GHz [1]. This imaging modality is possible due to the distinct dielectric properties of biological tissues within the microwave spectrum. The differentiation and imaging of tissues rely on their varying dielectric properties, particularly the significant contrast between tissues with high water content (such as tumor) and those with low water content (such as fat or bone) [2–4]. MWI in medical diagnostics offers notable advantages, particularly in breast and head imaging. Its appeal lies in cost-effectiveness, speed, non-ionizing characteristics, portability, and noninvasiveness. In breast imaging, MWI further stands out for its painless and less discomfort-inducing nature compared to conventional imaging techniques. Despite these benefits, the progress of MWI techniques has been delayed by hardware requirements and data quality. Encouragingly, recent advancements in wireless communication and computing technologies have paved the way for continued research and development in this field [4,5]. MWI can be divided into two major types, microwave tomography (MWT) and radar MWI (rMWI). In both approaches, a set of antennas transmits low-power microwave signals into the body, and in turn, these antennas measure the scattered microwave signals [6]. MWT is a quantitative technique aimed at recovering the dielectric properties of an object from scattered signals. This methodology is mainly used with narrowband signals. The resulting image is obtained using iterative image reconstruction algorithms to solve a non-linear inverse scattering problem. This approach requires regularization with the aim of achieving convergence to a meaningful solution since they are affected by the ill-posed problem [6]. rMWI, also referred to as ultra-wideband (UWB) radar or Confocal MWI, is a qualitative technique that relies on the contrast of dielectric properties of various tissues, for instance the contrast between healthy and malignant tissue for the detection of cancerous tissue. The goal is to build an image that shows dielectric © 2025 The Author(s). Published by IOP Publishing Ltd Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al variations without attempting to reconstruct the complete profile of dielectric properties [6]. The first algorithm for radar-based breast cancer detection was proposed by Hagness et al in 1998 [7,8] and since then several algorithms have been proposed. Over recent years, there has been a significant increase in the use of machine learning (ML) techniques within the field of MWI. ML, as an interdisciplinary area of artificial intelligence, focuses on developing algorithms capable of learning specific tasks across diverse research domains. In the context of MWI, ML has been applied to various tasks including image reconstruction, segmentation, and classification. While several literature reviews have examined this topic, their scope and depth vary considerably. Patel et al (2022) [9] addresses only reconstruction tasks, discussing the application of ML at each stage of the reconstruction process, but limits its examples to breast imaging and does not provide a thorough comparison of different possible applications. Yago et al (2023) [10] focuses exclusively on deep learning applications, emphasizing the challenges associated with these approaches and presenting only two practical study examples. Khalid et al (2024) [11], while offering a more detailed article-by-article analysis, is already outdated since the most recent study considered was published in 2023; moreover, it lacks the range of coverage presented in this scoping review. Shao (2025) [12] is also less comprehensive, addressing both classification and reconstruction tasks with general details. In contrast, the our scoping review offers a more detailed analysis by not only examining the algorithms employed but also considering key aspects of the datasets, including size and the characteristics of the devices used for data acquisition. Moreover, our scoping review evaluates reported results and performance metrics given the specific tasks addressed in each study, offering a clearer assessment of methodological effectiveness and practical relevance. This comparative evaluation reveals key trends, limitations, and research gaps that have been previously overlooked or addressed only superficially, thereby underscoring the significance and contribution of this work. This paper reviews the state of the art in the application of ML algorithms in MWI, with a focus on both image reconstruction and classification tasks, with either images or signals. The studies included in this comprehensive literature search were selected from three databases: Scopus, IEEE Xplore, and PubMed. The search strategy combined four thematic blocks using AND operators, and within each block, keywords were combined using OR operators: •Microwave, UWB, MWI, ultra-wideband, ultra wideband •Imaging, tomography, classification, detection, radar •ML, deep learning, artificial intelligence, ML, deep learning (DL), AI, neural network, convolutional neural network (CNN), artificial neural network (ANN), U-Net, support vector machine (SVM), k-nearest neighbor (KNN), decision tree (DT), random forest (RF), SVM, DT, KNN •Breast, brain, medical, diagnosis, diagnostics, disease, cancer, healthy, stroke This search returned a total of 1200 records-769 from Scopus, 302 from IEEE Xplore, and 129 from PubMed. After removing 372 duplicates, 828 records remained for screening. Of these, 717 were excluded based on the following inclusion criteria: •The study must involve MWI techniques as a central component. •The ML algorithms must be applied to image reconstruction and/or be used for classification tasks. •Classification tasks must involve clinically relevant classes, such as distinguishing between healthy and pathological conditions. •The article must present experimental, simulated, or clinical results, rather than being purely theoretical or conceptual. Additionally, 13 records were included to establish a baseline for classification performance using non-ML algorithms. In total, this scoping review comprises 124 studies, and the selection process, which was inspired by [13], is depicted in figure 1. Figure 2presents the distribution of the reviewed papers by application area. Among them, 37 focused on image reconstruction-22 on breast imaging, 14 on brain imaging, and one on leg imaging (not represented in the figure)-while 87 focused on classification tasks-62 on breast imaging and 25 on brain imaging. Section 2reviews ML applications for image reconstruction, with section 2.1 dedicated to breast cancer imaging and section 2.2 covering brain-related diseases such as brain cancer, Alzheimer’s disease (AD), and stroke. Section 3discusses ML-based classification, with section 3.1 focusing on breast disease classification and section 3.2 addressing brain-related conditions. Sections 2.1,3.1, and 3.2 organize the reviewed studies by algorithm type, separating them into ML and DL subsections. In cases where a study employs both ML and DL algorithms, it is presented in the ML subsection. This does not happen with the section 2.2, since all studies used DL algorithms. 2 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al Figure 1. Flowchart of the study selection process. A total of 1200 records were initially retrieved from search engines, with 372 duplicates removed prior to screening. An additional 13 records were manually included. After screening 828 records, 717 were excluded based on inclusion criteria, resulting in 124 studies included in the scoping review. Figure 2. Overview of the reviewed papers categorized by their focus within machine learning applications in medical microwave imaging. The studies are divided into image reconstruction (37 papers)-including breast imaging (22) and brain imaging (14)-and classification (87 papers)-including breast imaging (62) and brain imaging (25). Section 4critically examines the reviewed studies, outlining current limitations, challenges, and opportunities in applying ML to MWI. It highlights the transition from classical ML models to deep learning approaches and emphasizes the need for larger, more diverse datasets, improved generalization, robust evaluation metrics, and greater replicability through public datasets. Finally, section 5summarizes the key findings of the scoping review, suggesting the potential of ML to enhance MWI-based diagnostics and suggesting future directions in clinical validation, dataset expansion, and the development of more robust, real-world-ready models. 2. Reconstruction using ML models The use of ML algorithms to aid MWI reconstruction has been ongoing for over two decades, with one of the first studies being that of Rekanos et al [14], which present one of the first studies where they employ radial basis function neural network (RBFNN) to estimate the scatterer properties of tissues. Although their work did not focus on either brain or breast diseases, it is included here for its relevance. The RBFNNs were trained using the orthogonal least-squares algorithm, ensuring optimal network structure construction and straightforward determination of free parameters. The proposed methodology was employed in the estimation of the position and size of the proliferated marrow inside the bone of the lower part of the leg. The authors assumed the leg as a cylinder with a circular cross-section with a radius of 0.05 m and that the leg was irradiated from Nantennas, and the scattered field was measured from Lantennas. With the measurement configuration N×L, three different measurement configurations were tested: 20 ×20, 3 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al Table 1. Overview of image reconstruction papers in the state of the art (Target zone: breast). Reference Algorithm Microwave image type Type of data Kerhet et al [15] SVM Radar Numerical Ashtari et al [16] NNRGA Radar Numerical Conceiç˜ ao et al [17] SVM Radar Numerical Shah et al [18] CNN Tomography Numerical Khoshdel et al [19] CNN Tomography Numerical Khoshdel et al [20] CNN Tomography Numerical Mojabi et al [21] CNN Tomography Numerical Ambrosiano et al [22] ANN Tomography Numerical Mojabi et al [23] CNN Tomography Numerical Ambrosiano et al [24] ANN Tomography Numerical Ambrosiano et al [25] ANN Tomography Numerical Costanzo et al [26] U-Net Tomography Numerical Costanzo et al [27] U-Net Tomography Numerical No¨ el et al [28] PG-SACC-CNN Tomography Numerical Qin et al [29] CNN Tomography Numerical Fontaine et al [30] CNN Radar Experimental Bicer [31] CV-MWINet, RV-MWINet Radar Experimental Costanzo et al [32] CVNN Tomography Numerical Khoshdel et al [33] CNN Tomography Numerical Borghouts et al [34] CNN Tomography Numerical Franceschini et al [35] ANN Tomography Numerical Ambrosanio et al [36] ANN Tomography Numerical Artificial neural network (ANN), complex-valued combined model (CV-MWINet), complex-valued neural network (CVNN), convolutional neural network (CNN), neural network real genetic algorithm (NNRGA), physics-guided structurally-aware complex cascaded convolutional neural network (PG-SACC-CNN), real-valued combined model (RV-MWINet), support vector machines (SVM). 25 ×25, and 30 ×30. Gaussian noise was also added, with signal-to-noise ratios (SNRs) of 40, 35, 30, 25, and 20 dB, along with a noiseless case. In each of these datasets, the total number of samples, which in this study were the scattered-field measurements, is equal to 400. The study’s results showcase the RBFNNs’ precision in estimating the geometric properties of proliferated marrow. The best measurement configuration for low-noise or noiseless signals was 25 ×25, while for cases with high levels of noise, the best configuration was 30 ×30. The authors concluded that applying RBFNNs to medical imaging tasks, such as tissue characteristic estimation, holds promise due to their rapid estimation capabilities and resilience to noise. The study emphasized the importance of selecting an appropriate measurement configuration to balance information acquisition and noise impact. 2.1. Breast Table 1summarizes the papers on the application of ML algorithms to reconstruct breast images using MWI data. 2.1.1. ML Kerhet et al [15] proposed a 3D approach using a SVMs classifier with a Gaussian kernel, to obtain the probability maps of tumor presence in a phantom of the breast. The approach was evaluated on synthetic data generated with the finite element method (FEM). The dataset used in this study contains 41 250 samples, each comprising 99 features: 96 time-steps of the signal and 3 geometrical coordinates of each pixel. Each sample is labeled with a binary class, where the positive class indicates a tumor-affected area, and the negative class represents a tumor-free area. There were 28 750 samples for the training set, 8750 samples for the validation set, and 3750 samples for the test set. All sets were scaled to a range from −1 to 1. The authors tested the approach in two different scenarios: one where the dielectric properties of the breast tissue were known a priori, and another where these properties were unknown. Additionally, the authors attempted training on a reduced dataset for the noiseless case of the first scenario. They reported that the probability maps obtained for the first scenario showed that the region near the tumor location tended to clearly stand out against the background. For the second scenario, they observed that although the quality of results was lower, the highest probability values corresponded to the tumor voxels. The specific method used to evaluate the model’s performance was not reported, nor were objective results presented, making it unfeasible to ensure reproducibility or to draw conclusions supported by the model’s performance. 4 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al Conceiç˜ ao et al (2017) [17] used an SVM to help in the diagnosis of breast cancer, while using a rMWI prototype developed at the University of Bristol. This study extends the authors’ previous work [37]. In this study, 36 numerical phantoms were used, developed with layers of tissue mimicking materials. The model considered in the study consisted of 1558 synthetic focal points (SFPs), which corresponded to voxels in these numerical phantoms, comprising a total of 1925 455 measurements, classified as either ‘hits’ (tumor tissue) or ‘misses’ (healthy tissue) in the 36 numerical phantoms. Of the SFPs, 779 were ‘hits’ and 779 were ‘misses.’ An SVM classifier was trained using 24 features extracted from the signal of each SFP, aiming to distinguish between ‘hits’ and ‘misses.’ The dataset was divided into training and testing sets, and a k-fold crossvalidation with k=10 was implemented. The authors concluded that the SVM classification approach outperformed the linear discriminant analysis (LDA) classifier presented in their previous work [37]. 2.1.2. Deep learning Ashtari et al (2010) [16] proposed a neural network real genetic algorithm (NNRGA) for breast cancer detection using MWI. The dataset comprised 200 numerical breast profiles and 200 random profiles, which featured randomly distributed tissue types. The data were divided into training, validation, and test sets using a 60%/20%/20% split. Four breast tissue types were used in the study: mostly fatty, scattered fibroglandular, heterogeneously dense, and very dense. The NNRGA achieved its best permittivity reconstruction performance with the scattered fibroglandular breast type, reaching a relative error of just 2.7%. Shah et al (2018) [18] employed a CNN to estimate the total electric field from the Born-approximated electric field, treating the problem as an image-to-image transformation. To train the network, the authors used a dataset composed of nine magnetic resonance imaging (MRI)-derived breast numerical phantoms from a public database [38], generating corresponding permittivity and conductivity maps. From these, 75% of the 2D coronal slices from seven randomly selected numerical phantoms were used for training, 25% for validation, and the remaining two numerical phantoms were reserved for testing. The images were normalized and had a resolution of 0.5 mm, while the simulated MWI images had a lower resolution of 1.5 mm. Training was performed using 33 ×33 image patches. This learning-based approach significantly reduced the l2distance between the predicted and true contrast distributions from 57.75 to 28, indicating its effectiveness in capturing the nonlinearities of MWI. The authors concluded that estimating the total electric field from the Born-approximated field using a CNN is a feasible and promising strategy for improving reconstruction accuracy in MWI. Khoshdel et al (2019) [19] proposed a deep learning approach based on CNN to enhance MWI reconstructions in a dual-modality microwave-ultrasound system. The study employed 1200 numerically simulated breast phantoms, with half containing one tumor and the other half containing two. A hybrid reconstruction strategy was introduced, where the output of the contrast source inversion (CSI) method-augmented with ultrasound-derived tissue profiles-was used as input to the CNN. The CNN significantly outperformed CSI in both root mean square error (RMSE) and area under the ROC curve (AUC), achieving an RMSE of 0.122 and AUC of 0.987, compared to CSI’s RMSE of 2.199 and AUC of 0.897. In a subsequent study, Khoshdel et al (2020) [20] extended this work to 3D MWI using a 10-channel 3D U-Net architecture trained on synthetic data. The dataset comprised 600 numerical breast phantoms with tumor diameters ranging from 1.1 to 1.5 cm, equally divided between singleand double-tumor cases. The CNN was trained on CSI reconstructions obtained using a FEM-based inversion and evaluated on both synthetic and experimental test sets. On synthetic data, the CNN achieved an AUC of 0.957 and RMSE of 1.161, while on experimental data, it obtained an AUC of 0.938 and RMSE of 1.172. In both cases, the CNN outperformed CSI, which achieved AUCs of 0.935 (synthetic) and 0.794 (experimental), and RMSEs of 1.436 and 1.250, respectively. Mojabi et al (2020) [21] introduced a CNN with U-Net architecture to reconstruct MWI images from quantitative dielectric and/or ultrasonic property maps. The dataset included 400 numerical breast phantoms: 150 derived from 3D MRI intensity models capturing diverse fibroglandular patterns, and 250 from 2D MRI intensity models. Among these, 50 phantoms contained tumors across eight distinct configurations. The CNN was trained to map quantitative reconstructions of dielectric and/or ultrasonic properties to corresponding tissue-type and uncertainty maps. Although dataset partitioning was not specified, results showed that the CNN outperformed a Bayesian model in terms of the number of correctly classified pixels, highlighting improved accuracy in tissue classification. In follow-up work, Mojabi et al (2021) [23] employed two U-Net-based CNNs-one with two depth levels and the other with four-to predict the complex nonlinear relationship between ultrasound compressibility and dielectric properties. The two-level U-Net used 400 numerical phantoms derived from eight tumor-free, MRI-based breast models representing fatty, dense, and heterogeneous tissue types (350 for training, 50 for testing). The four-level U-Net dataset was augmented with 11 rotations at 30◦intervals, expanding it to 4800 phantoms (4200 training, 600 testing). Inputs were ultrasound compressibility images reconstructed under the Born 5 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al approximation, and outputs were the real and imaginary parts of the true relative complex permittivity. The CNNs yielded predictions closely matching the true dielectric properties in both shape and magnitude, demonstrating their potential for generating accurate prior information for MWI algorithms. However, the study noted that training under the Born approximation may lead to the omission of tumor regions in some reconstructions. More recently, Khoshdel et al (2023) [33] proposed a multi-branch CNN architecture to reconstruct two-dimensional permittivity maps for breast cancer detection using MWI. The model comprises two input branches-one for background tissue information and another for scattered field data-and was trained on 4800 synthetic 2D images generated from MRI-based numerical breast phantoms developed in earlier studies [21,23]. While the data split was not disclosed, the test set included tumor-free, single-tumor, and two-tumor cases. Reconstruction quality was assessed through qualitative visual analysis, with the authors reporting that the CNN was effective in addressing the electromagnetic inverse scattering problem in breast MWI. Ambrosanio et al (2020) [22] investigated the application of artificial neural network (ANN) for MWI of numerical breast phantoms containing tumors. A dataset of 50 000 numerically simulated breast profiles was employed, divided into training (80%), testing (15%), and validation (5%) sets. Although no quantitative metrics were reported, qualitative visual analysis revealed that the ANN-based reconstructions yielded superior tumor localization and contrast recovery compared to the traditional distorted born iterative method (DBIM). Building upon this work, Ambrosanio et al (2022) [24] explored ANN-based quantitative breast MWI using a dataset of 120 000 simulated profiles categorized into four breast tissue classes (A to D), ranging from predominantly adipose (Class A) to highly fibroglandular tissue (Class D). Data were split into 85% for training, 10% for validation, and 5% for testing. The ANN was evaluated against two conventional techniques, AMTISTA and CC-CSI, using structural similarity index measure (SSIM), normalized RMSE (NRMSE), and correlation coefficient (CORR) for both permittivity and conductivity reconstructions. The ANN outperformed both methods across all metrics and tissue classes for permittivity, and in most cases for conductivity, except for SSIM in Classes B and D, where CC-CSI slightly surpassed the ANN. The best ANN results for conductivity reconstruction were in Class D (SSIM: 0.417, NRMSE: 0.087, CORR: 0.871), while the highest SSIM for permittivity was in Class A (0.458), the lowest NRMSE in Classes A and B (0.102), and the highest CORR in Classes B and C (0.818). These outcomes highlight the ANN’s capability to produce high-fidelity reconstructions across varying tissue compositions. A complementary study by Ambrosanio et al (2022) [25] employed a fully connected ANN trained on the same dataset (split 85%/10%/5%) to reconstruct both permittivity and conductivity, achieving NRMSE values of 0.10 and 0.26, respectively. Borghouts et al (2023) [34] applied a CNN to image reconstruction and classification of breast MWI data into malignant or healthy categories. The dataset consisted of 160 000 simulated breast profiles, divided into 128 000 for training, and 16 000 for validation and 16 000 for testing, balanced across both classes. Reconstruction performance was assessed using Soft-Dice, normalized cross-correlation (NCC), and NRMSE, while classification performance was evaluated with accuracy, sensitivity, specificity, precision, and F1-score. The CNN achieved a Soft-Dice score of 0.144, NCC of 0.337, and NRMSE of 0.809. For classification, it demonstrated excellent performance, achieving 99.95% accuracy, 99.96% sensitivity, 99.94% specificity, 99.94% precision, and a 99.95% F1-score, confirming the potential of deep learning for end-to-end breast cancer detection in MWI. Using the same dataset, Franceschini et al (2023) [35] implemented an ANN for breast cancer detection, evaluating reconstruction performance on a pixel-wise basis resulting in accuracy (99.5%), sensitivity (98.9%), specificity (99.9%), and AUC (100%). Despite employing the same dataset as Borghouts et al, direct comparisons are limited due to differences in methodologies. Extending their work, Ambrosanio et al (2024) [36] developed a fully connected ANN to generate tumor probability maps from breast MWI data. Using the same dataset of 160 000 simulated breast profiles (80% training, 10% validation, 10% testing), the model achieved a Soft-Dice score of 0.102, NCC of 0.287, and NRMSE of 0.856, indicating reasonable performance and reinforcing the utility of ANNs for probabilistic tumor localization in MWI. Costanzo et al have published several studies utilizing data from the University of Wisconsin Computational Electromagnetics (UWCEM) repository [38], in which the output of the quadratic born iterative method (BIM) was used as input to ML models. In Costanzo et al (2022) [26], a CNN was employed; however, the total number of samples and the dataset partitioning were not fully disclosed. The CNN achieved a mean relative error of 8.09% and an average reconstruction accuracy of 91.91% across four different breast phantoms. In a complementary study from the same year, Costanzo et al (2022) [27] used a more controlled dataset consisting of 177 images from two numerical phantoms, with 95% of the images allocated for training and 5% for testing. The CNN achieved relative errors of 3.8% and 7.18% for lowand high-density phantoms, respectively. However, these results may reflect overfitting due to the small dataset size and limited phantom diversity. In subsequent work, Costanzo et al (2023) [32] introduced a 6 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al complex-valued neural network (CVNN) and compared its performance with two real-valued neural network (RVNN), one trained to reconstruct permittivity and the other conductivity. The dataset comprised 1500 images, divided into 80% for training and 20% for validation. The CVNN achieved a validation loss of 8.4205, closely matching the RVNN for permittivity (8.0074) and outperforming the RVNN for conductivity (16.2775), suggesting that the CVNN provides comparable or improved performance, particularly when handling complex-valued data. No¨ el et al (2022) [28] proposed a physics-guided structurally-aware complex cascaded CNN (PG-SACC-CNN) for breast MWI, integrating both microwave and ultrasound modalities. The performance of the PG-SACC-CNN was compared to that of various CNN architectures, including a ResNet-based model. All networks were trained and tested using breast phantom data from the UWCEM repository [38]. The dataset included 2920 samples, divided into 2336 for training and 584 for testing. Network performance was assessed using Intersection over Union (IoU), reconstruction error, and SSIM. The PG-SACC-CNN achieved superior results across all metrics, with an IoU of 0.755, a reconstruction error of 0.156, and an SSIM of 0.810, indicating enhanced reconstruction accuracy and preservation of structural features. Similarly, Qin et al (2022) [29] employed several CNN variants to jointly invert microwave and ultrasonic data for breast imaging, using the same UWCEM breast phantom dataset. In this case, the dataset comprised 2180 samples, with 1920 allocated for training and 240 for testing (there are 20 samples unaccounted for). Among the tested models, the CNN architecture that employed multistream input and multitask learning demonstrated the highest performance, achieving an IoU of 0.7504. These results further support the effectiveness of hybrid CNN-based strategies for multimodal breast imaging tasks. Fontaine et al (2023) [30] used a CNN in a portable microwave system, aimed to reduce the device’s cost, size, and complexity. The CNN was trained to directly reconstruct rod phantoms from their S11 sinograms. The training and testing data consisted of pairs of numerical rod phantoms and their corresponding S11 sinograms. In total, 10 000 phantoms were generated, but by randomly flipping and rotating the pairs, the dataset was expanded to 250 000 image-sinogram pairs. The entire dataset was then split into a training set (80%) and a test set (20%). The authors concluded that the CNN’s predictions were promising compared to Delay-And-Sum reconstructions and that similar devices, using inexpensive materials, could serve as a viable alternative to conventional image reconstruction methods. Bicer (2023) [31] evaluated four deep learning models for microwave radar-based breast imaging: a real-valued deep neural network (RV-DNN), a real-valued CNN (RV-CNN), a real-valued combined architecture (RV-MWINet), and a complex-valued combined model (CV-MWINet). The dataset consisted of 1000 scattered signals acquired from experimental breast phantoms. The RV-DNN and RV-CNN models were evaluated using mean squared error (MSE) and SSIM, while RV-MWINet and CV-MWINet were assessed using SSIM and classification accuracy. All models were evaluated with 10-fold cross-validation. The best performance was achieved by RV-MWINet and CV-MWINet, both attaining an SSIM of 0.999 and a classification accuracy of 99.3%. In comparison, the RV-DNN and RV-CNN models achieved SSIM values of 0.914 and 0.911, and MSE values of 197.401 and 162.089, respectively, demonstrating the higher performance of the combined architectures for both reconstruction and classification of breast tumor profiles. Table 2summarizes the characteristics of the acquisition systems employed in the studies discussed in this section. These include the number of antenna positions, the configuration of the acquisition system, the mobility of the antenna array, and the operating frequencies. The ‘No. of antenna positions’ column shows that most studies use setups with more than 15 antenna elements, commonly ranging from 20 to 30, with only a few exceptions reporting lower counts-such as the study by Kerhet et al (1 T and 16 R), Ashtari et al (4 T 16 R), and those by Costanzo et al, which used 10 or 18 antennas for both transmission and reflexion. The ‘Multistatic vs Monostatic’ column reveals a clear dominance of multistatic configurations, where each transmitted signal is received by all antennas in the array. Monostatic or quasi-multistatic (defined as the transmitted signal is received by all antennas except the transmitting one) systems are less common, and bistatic configurations appear rarely, as in Kerhet et al ‘s work. The ‘Fixed vs moving’ column highlights a strong incidence of fixed antenna systems, used in nearly all studies, with Bicer et al being the only example employing a moving array. The ‘Frequencies’ column indicates that 1 GHz is the most frequently used operating frequency across studies, although several works adopt broader or multiple frequency bands-such as Conceiç˜ ao et al (3–8 GHz), Fontaine et al (0.7–3 GHz), and Bicer et al (1–10 GHz)-to enhance imaging quality. Overall, the table reflects a consistent trend in the adoption of fixed, multistatic systems with moderate-to-high numbers of antenna elements and operating frequencies centered around 1 GHz. These choices suggest a growing consensus on hardware configurations that balance complexity, resolution, and practicality in breast MWI research. Table 3summarizes key characteristics of the datasets used in the reviewed studies. The ‘Dataset’ column describes both the dataset size and what the reported size refers to, which varies across works-for instance, it may represent the number of raw signal samples, image pairs, or patient profiles. This variability reflects the 7 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al Table 2. Summary table on the equipment used for microwave signal acquisition (Target Zone: Breast). Label ‘T’ represents transmitters, while ‘R’ represents receivers. When these labels are absent it means the respective study did not specify transmitters and receivers. The (∗) indicates that the nomenclature used in this paper differs from that employed by the authors of the paper. ‘NA’ means not available. Reference No. of antenna positions Monostatic vs multistatic Moving vs fixed Frequencies Kerhet et al [15] 1 T 16 R Bistatic Fixed 6 GHz Ashtari et al [16] 4 T 16 R Multistatic Fixed 1 GHz Conceiç˜ ao et al [17] 60 Multistatic Fixed 3–8 GHz Shah et al [18] 24 T 24 R Multistatic Fixed 2 GHz Khoshdel et al [19]NA Multistatic NA NA Khoshdel et al [20] 24 Multistatic Fixed 1.1–1.5 GHz Mojabi et al [21] 30 Multistatic Fixed 1, 1.5 and 2 GHz Ambrosiano et al [22] 30 Quasi-multistatic∗Fixed 1 GHz Mojabi et al [23] 30 Multistatic Fixed 1 GHz Ambrosiano et al [24] 30 Quasi-multistatic∗Fixed 0.2–1 GHz Ambrosiano et al [25] 30 Quasi-multistatic∗Fixed 1 GHz Costanzo et al [26] 18 T 18 R Multistatic Fixed 1 GHz Costanzo et al [27] 18 T 18 R Multistatic Fixed 1 GHz No¨ el et al [28] 20 Multistatic Fixed 1 GHz Qin et al [29] 20 Multistatic Fixed 1 GHz Fontaine et al [30] 26 Multistatic Fixed 0.7–3 GHz Bicer [31] 90 T 90 R Monostatic Moving 1–10 GHz Costanzo et al [32] NA Multistatic NA NA Khoshdel et al [33] 30 Multistatic Fixed 1, 1.5 and 2 GHz Borghouts et al [34] 30 Quasi-multistatic∗Fixed NA Franceschini et al [35] 30 Quasi-multistatic∗Fixed 1 GHz Ambrosanio et al [36] 30 Quasi-multistatic∗Fixed 1 GHz lack of standardization in how dataset size is reported. The ‘Input sample’ column specifies the actual data type provided to the model, such as scattered fields, dielectric profiles, or reconstructed images. The distinction between the ‘Dataset’ and ‘Input sample’ columns is necessary because most studies do not clearly define how many input samples are derived from the dataset. To address this, the table descriminates general dataset size from specific model inputs. The ‘Metric’ column reports the performance metric used and its corresponding value, with notable variation across studies. While some consistency exists, MSE and SSIM are frequently reported, other metrics such as IoU, AUC, accuracy, and F1-score also appear. The diversity in data types, input formats, and evaluation criteria highlights the heterogeneous nature of research in this domain. However, more recent studies tend to use larger datasets, more comprehensive performance metrics, and achieve higher accuracy or reconstruction quality, indicating a trend toward increasingly robust and data-driven approaches. This section reveals a clear predominance of DL algorithms in breast image reconstruction tasks. Although synthetic (numerical) data remains more commonly used, several studies employing experimental data-such as phantom-based measurements-have reported promising results. Additionally, when comparisons are possible using the same evaluation metrics, there is a visible trend of improved performance as dataset sizes increase. These encouraging outcomes highlight the importance of continued development in this field. Advancing these applications will require the creation of larger and more diverse datasets, especially incorporating clinical data, to ensure that models are trained with information reflective of real-world conditions. 2.2. Brain The articles that combined ML and MWI for the reconstruction of brain images associated with brain cancer, stroke and AD are summarized in table 4. Xiao et al (2022) [39] introduced the hybrid neural network electromagnetic inversion scheme (HNNEMIS) for super-resolution 3D MWI of the brain. The proposed framework integrates shallow and DNN architectures by combining a semi-joint backpropagation neural network (SJ-BPNN) with a U-Net CNN. Two numerical scenarios were evaluated: one that involved normal brain imaging with a voxel resolution of 256 ×256 ×256, and another focused on abnormal scatterer detection with a resolution of 512 ×512 ×512. Each case used a training dataset of 200 samples and was evaluated under both noise-free and noisy conditions. The model achieved a reconstruction (model) misfit of 12.52%, and a data misfit of 0.73%, demonstrating its potential for high-resolution brain imaging. 8 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al Table 3. Summary of dataset characteristics and evaluation metrics from studies listed in table 1. Reference Dataset Input sample Metric Kerhet et al [15] 41 250 (Signals and coordinates of the cell) Signals and coordinates of the cell NA Ashtari et al [16] 200+200 (Breasts +random profiles) FV Relative error: 2.7% Conceiç˜ ao et al [17] 1558 (SFPs) Signal NA Shah et al [18] NA NA L2 Distance: 28.54 Khoshdel et al [19] 1200 (Breast phantoms) CSI reconstruction RMSE, AUC: 0.122, 0.987 Khoshdel et al [20] 600 (Training set, breast phantoms) CSI Reconstructions RMSE, AUC (Syntetic, experimental): 1.161, 0.957; 1.172, 0.938 Mojabi et al [21] 400 (Breast phantoms) Dieletric and/or ultrasound reconstructions Overall error: 0.024 Ambrosiano et al [22] 50 000 (Breasts profiles) Breast profiles NA Mojabi et al [23] 400, 4800 (Breast phantoms) Reconstructed ultrasonic (Breast Compressibility) MSE: 0.097, 0.038 Ambrosiano et al [24] 120 000 (Breast profiles) Breast profiles NRMSE, SSIM, CORR: (0.102, 0.458, 0.795 Ambrosiano et al [25] 120 000 (Breast profiles) Breast profiles NRMSE (Permittivity, conductivity): 0.10, 0.26 Costanzo et al [26] NA Quadratic BIM output Acc, RE: 91.91%, 8.09% Costanzo et al [27] 177 (BIM images) Quadratic BIM output Relative error: 3.87% No¨ el et al [28] 2920 (Samples) Scattered field IoU, Error, SSIM: 0.755, 0.156, 0.810 Qin et al [29] 2180 (Samples) Scattered field IoU: 0.7504 Fontaine et al [30] 250 000 (Sinogram-image pairs) Sinogram-image pairs MSE, Accuracy: 0.137, 63% Bicer [31] 1000 (Scattered fields) Scattered field Acc, SSIM: 99.3%, 99.9% Costanzo et al [32] 1500 (Contrast maps) Quadratic BIM output Validation loss: 8.4205 Khoshdel et al [33] 4800 (Breast phantoms) Scattered field NA Borghouts et al [34] 160 000 (Breast profiles) Scattered field Acc, Sen, Spe, Prec, F1: 99.95%, 99.96%, 99.94%, 99.94%, 99.95% Franceschini et al [35] 160 000 (Breast profiles) Scattered field Acc, Sen, Spe, AUC: 99.5%, 98.9%, 99.9%, 100% Ambrosanio et al [36] 160 000 (Breast profiles) Breast profiles NRMSE, NCC, Soft-DICE: 0.856, 0.287, 0.102 Accuracy (Acc), area under the curve (AUC), breast imaging modality (BIM), contrast-source inversion (CSI), correlation coefficient (CORR), F1-score (F1), Intersection over Union (IoU), mean squared error (MSE), not available (NA), normalized cross-correlation (NCC), normalized root mean square error (NRMSE), precision (Prec), relative error (RE), sensitivity (Sen), specificity (Spe), structural similarity index measure (SSIM). From the same research group, Cheng et al (2022) [40] proposed a 3D Full Convolution Electromagnetic Reconstruction Neural Network (3D-FCERNN), supplemented by a U-Net-based image enhancement module. The 3D-FCERNN directly maps scattered field measurements to three-dimensional distributions of relative permittivity and conductivity, while the U-Net module refines the outputs to enhance image quality. Two cases were studied using simulated brain phantom data: the first with 190 healthy numerical phantoms (256 ×256 ×256 resolution) and the second with 250 unhealthy numerical phantoms (512 ×512 ×512 resolution). Both models were tested under various levels of Gaussian noise. The inclusion of the U-Net significantly improved imaging quality, particularly under high noise conditions (−10 to −40 dB). At −40 dB, the enhanced model achieved a 19.36% model misfit and 15.37% data misfit, outperforming the FCERNN alone. Zhao et al (2022) [41] developed a ML-based inversion approach with resolution enhancement, incorporating a SJ-BPNN, a U-Net, and modified akima piecewise cubic hermite interpolation. The method was trained on 10 620 scattered field measurements derived from one numerical brain phantom. The resulting model achieved a 13.02% model misfit and a 5.7% data misfit, indicating effective reconstruction accuracy. Xue et al (2024) [46] implemented a modified U-Net architecture, U-Net3++, for brain image reconstruction based on MWI. The study also compared the performance of alternative CNN architectures, 9 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al Sacristán et al (2016) [78] employed KNN and SVM models to classify simulated breast models as healthy or tumorous. Two distinct datasets were used: the first, referred to as the ‘Homogeneous’ dataset, included 1534 numerical breast models composed solely of skin and fatty tissue; the second, the ‘Heterogeneous’ dataset, consisted of 2395 breast numerical models containing additional fibro-glandular patches, with the fibro-glandular to fatty tissue ratio varying between 0% and 25%. In both datasets, 50% of the numerical models contained a tumor. The SVM models outperformed the KNN models in both cases, achieving 94% vs 73% accuracy, respectively, on the Homogeneous dataset, and 62% vs 57% on the Heterogeneous dataset. Oliveira et al (2018) [81] applied a diagnostic architecture consisting of steps to create breast and tumor models, acquiring signals via FDTD simulation, followed by data processing, and finally, diagnosis. The breast models were derived from the repository created by the UWCEM laboratory [119]. The study used three heterogeneous breast models, with the percentage of glandular tissue ranging from 1% to 27%. For tumor models, the authors employed an algorithm from [120] to generate 72 unique tumor models, categorized into two classes: smooth borders representing BTs, and spiculated borders representing MTs. With a microwave device, 1080 microwave scans were performed, where each scan is composed of backscattered signals collected from 78 independent channels. The dataset used in this study comprised a total of 84 240 signals. Two methods, tumor windowing (TW) and FE, were used for signal processing. The classification was performed using RFs, applied to three types of classification models. The authors implemented a validation methodology based on nested cross-validation, and the performance of the classification model was assessed by plotting ROC curves. The AUC of the ROC Curve was also used to evaluate the model, as a higher AUC indicates a more generalizable model. The authors concluded that the models performed better when handling signals collected under consistent conditions. Furthermore, the data pre-processed solely with FE slightly outperformed those processed with TW +FE or only TW. Aydin et al (2019) [82] applied a RF algorithm for breast cancer detection using data generated from a numerical hemispherical breast phantom. The dataset consisted of 400 samples, incorporating four different experimental conditions: with both matching liquid and tumor, with matching liquid but no tumor, with tumor but no matching liquid, and with neither. However, the paper did not disclose the distribution of samples across these categories. Model performance was evaluated using 10-fold cross-validation, yielding a mean classification accuracy of 94%. Aldhaeebi et al (2019) [83] proposed a breast cancer detection method using ML algorithms, specifically SVM and DT classifiers. The dataset consisted of 90 numerical breast phantoms, evenly divided between H and NH cases. Among the models evaluated, the DT classifier achieved the best performance, with 65% accuracy, 68.8% precision, 55% recall, and 75% specificity. MammoWave is a microwave device used for MWI acquisition and has supported several ML-based breast classification studies. Early work by Rana et al (2019) [84] used data from 18 subjects (12 H and 11 NH breasts) and tested KNN, SVM, and MLP classifiers with varying training set sizes. The best performance was achieved by SVM using 40% of the data, with 98.9% accuracy, 97.7% sensitivity, 99.7% specificity, and an Matthew’s CORR (MCC) of 0.955. In a later study, Rana et al (2021) [87] applied SVM to 61 breast scans (25 H, 36 NH), achieving 91% accuracy, 84.4% sensitivity, and 97.2% specificity. Building on these efforts, Papini et al (2023) [97] conducted a more extensive clinical study on 697 breasts (123 NH, 574 H) using multiple classifiers (SVM, RF, DT, KNN, etc) with dimensionality reduction and class-balancing techniques. SVM and RF delivered the best results depending on the subset of data used: SVM achieved 88% accuracy, 86% sensitivity, and 89% specificity using raw data; RF reached 85% accuracy, 83% sensitivity, and 90% specificity for dense breasts; and SVM again showed the following results using image features (up to 87% accuracy and 88% specificity). Ghavami et al (2023) [101] reported on the same dataset with 697 samples, using SVM to distinguish malignant from benign or tumor-free breasts, achieving 90% accuracy, 90% sensitivity, and 92% specificity. Rana et al (2023) [102] used data from 61 breasts (35 malignant), with SVM again yielding improved results: 95.5% accuracy, 97.2% sensitivity, 94.5% specificity, and an MCC of 0.909. In contrast, Shadwell et al (2024) [106] reported mixed outcomes: SVM reached 80% accuracy on a dataset of 352 breasts but dropped significantly to 44% on a separate set of 217 samples. More recently, Taghipour-Gorjikolaie et al (2024) [110] an SVM achieved 64.49% accuracy, 61.25% sensitivity, and 65.08% specificity, with a larger dataset of 1024 breasts (161 non-healthy, 863 healthy). Overall, SVM has been the most widely used and successful algorithm across studies, consistently outperforming other classifiers, especially on smaller and more balanced datasets. However, performance tends to degrade on larger, unbalanced datasets, highlighting the need for robust data preprocessing and model selection. Reimer et al (2019) [85] evaluated tumor detection performance using two ML models: SVM and MLP. Simulated data were generated from 2D numerical breast phantoms based on BI-RADS Class 1 and Class 2 tissue densities. Two datasets were used, comprising 2000 phantoms for Class 1 dataset and 2001 for Class 2 16 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al dataset, with 50% of each dataset containing malignant tissue. For both classes, two sub-datasets were created: one using data from a single rotational position (α) and another using data from five positions (β), resulting in different feature dimensionalities. The classifiers were trained using five preprocessing pipelines, and their performance was evaluated using accuracy, sensitivity, specificity, and the AUC. The best results were obtained with the (β) dataset. For Class 1, the SVM achieved 88% accuracy, 84% sensitivity, 92% specificity, and a ROC AUC of 95%, while for Class 2 it achieved 87%, 83%, 91%, and 94%, respectively. The MLP achieved similar results: for Class 1, 88% accuracy, 86% sensitivity, 90% specificity, and 94% ROC AUC; and for Class 2, 85% accuracy, 83% sensitivity, 86% specificity, and 92% ROC AUC. Across all classifier-dataset combinations, the feature scaling pipeline consistently yielded the best performance. Fasoula et al (2021) [88] conducted a clinical investigation involving 25 symptomatic patients to explore the capability of the Wavelia semi-automated quantitative imaging function (QIF) in detecting and classifying breast lesions as malignant or benign. This study presents the methodology used in the Wavelia QIF to support breast lesion detection based on lesion persistence, sizing, and characterization within a feature space that encompasses both shape-based and texture-based features. The low-dimensional feature space covered shape-based features such as solidity, and texture-based features such as correlation and busyness. The features were extracted from the images. In the Wavelia QIF, a NB and QDA classifier were trained for the purpose of distinguishing between malignant and benign breast lesions. The results showed a notable level of separability between malignant and benign breast lesions, with a classification loss of 11.5% estimated using 10-fold cross-validation for the trained QDA classifier. Chen et al (2021) [89] implemented SVM, DT and RF classifiers to distinguish between benign and malignant breast tumors using a two-dimensional numerical breast model. A total of 600 samples were generated, although the distribution between the two classes was not disclosed. The models were evaluated using both accuracy and F1-score. Among the three classifiers, the SVM model achieved the best performance, with an accuracy of 72.5% and an F1-score of 72.8%. Liu et al (2021) [90] proposed a FE method based on ensemble empirical mode decomposition (EEMD) for breast tumor detection, evaluated using an SVM classifier. The numerical breast models were derived from four MRI images, yielding a dataset of 11 232 backscatter signal sets, with a balanced class distribution (healthy vs non-healthy) in a 1:1 ratio. The EEMD method involved iteratively adding white noise to the signal and applying empirical mode decomposition, resulting in several representative intrinsic mode function (IMF) components. FE was then applied to both the original signals and the IMFs, followed by dimensionality reduction using PCA before training and testing the classifier. The SVM achieved an accuracy of 84.5%. However, it is not possible to draw meaningful conclusions regarding the effectiveness of the EEMD-based approach, as no comparison with alternative FE methods was provided. Martins et al (2021) [93] investigated the impact of antenna design on breast tumor detection performance using ML algorithms, specifically KNN, SVM, and LDA. Two antennas were tested: a Vivaldi antenna and a slot-based antenna. The dataset consisted of 480 measurements obtained from five experimental breast phantoms-240 without any tumor and 240 with one of two different tumor embedded. For the slot-based antenna, classification accuracies were 60% for KNN, and 50% for both SVM and LDA. In contrast, the Vivaldi antenna yielded significantly higher performance, with 80% accuracy for KNN, 70% for SVM, and 85% for LDA, demonstrating a clear advantage of the Vivaldi design over the slot-based alternative, which indicates that antenna design may have an impact on classification results. Reimer et al (2020) [121] developed an open-source dataset for breast microwave MWI, which has been widely adopted for evaluating ML approaches to breast tumor detection. The dataset is divided into three generations: the first comprises 249 experimental scans, the second includes 1008 scans, and the third contains 200 scans. While the first two generations were reported in the original publication [121], a third generation was introduced later by Reimer et al (2021) [122]. Sami et al (2021) [91] used the second-generation dataset of 1008 scans-800 for training and 208 for cross-validation-to train an SVM model, achieving outstanding results: 99.7% accuracy, 99.2% sensitivity, and 99.9% specificity. Patel et al (2021) [92] employed a subset of 249 scans to compare DT, RF, and XGBoost models. The RF model achieved the best performance among these, with 94% accuracy, 100% sensitivity, and 89.2% specificity; however, none of the models matched the performance of the SVM reported by Sami et al (2021) [91]. Dridi et al (2024) [107] evaluated linear and RBF-kernel SVMs, as well as logistic regression (LR), on a subset of 249 scans (199 for training and 50 for testing). The linear SVM achieved the best results in this study, with 98% accuracy, 96.15% sensitivity, and 100% specificity. Reimer et al (2025) [114] employed the third-generation dataset of 200 scans (150 for training and 50 for testing) to train LR, SVM, and RF models using features extracted from reconstructed images. Although multiple combinations of classifiers and reconstruction methods were tested, the study did not report 17 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al specific AUC values. In summary, using accuracy as the comparison metric, the best results were obtained in studies that used a single dataset generation, with the highest performance achieved by Sami et al [91] using the second generation, which likely corresponds to a real case scenario where data variability is present. The study based on the third generation, reported by Reimer et al [114], did not disclose objective performance metrics, making direct comparisons with the other generations infeasible. Another microwave device used for MWI acquisition and ML-based classification is the Scan and Find Early (SAFE) system, which has been employed in several clinical studies. Janjic et al (2022) [94] first evaluated the device using backscattered signals from 54 patients with breast lesions (29 benign and 25 malignant), applying a gradient boosting algorithm that achieved 81% accuracy, 80% sensitivity, and 83% specificity. Building on this, Janjic et al (2023) [96] expanded the dataset to 113 samples (70 benign, 43 malignant), collected from BI-RADS 4 and 5 patients, and used an AdaBoost classifier, which resulted in slightly lower performance: 78% accuracy, 79% sensitivity, and 77% specificity. In a more extensive clinical study, Janjic et al (2023) [98] analyzed 558 breast samples (390 from healthy subjects and 168 from cancer patients) using an SVM classifier, reporting significantly improved outcomes with 93% accuracy, 95% sensitivity, and 92% specificity. Similarly, Yurtseven et al (2025) [113] assessed the SAFE device on 526 cases (286 non-cancerous and 240 malignant), employing an XGBoost model that achieved its best performance on dense breast tissue, with 91% accuracy, 91% sensitivity, and 92% specificity. Collectively, these studies highlight the SAFE device’s growing clinical relevance, with progressively larger datasets and more advanced classifiers yielding higher diagnostic performance. Abdelmawgood et al (2023) [103] proposed a tumor detection framework based on textile antenna sensors and ML algorithms, including LR, SVM, NB, DT, XGBoost, AdaBoost, CatBoost, and ANN. The dataset was generated using a three-layer numerical breast phantom consisting of fat, fibroglandular, and skin tissues. It comprised 1001 samples, balanced between healthy and malignant cases, although the class distribution was not explicitly detailed. The dataset was split into training and test sets using a 70/30 ratio. Results were reported for two tumor configurations: two centered tumors and one shifted tumor. However, it was not clearly stated whether 1001 samples were used per configuration or in total. The best performance was achieved by XGBoost for the 5 mm centered tumor case, with an accuracy of 99%. The worst performance was observed with NB, which achieved only 49% accuracy for both the 5 mm centered and shifted tumor cases. These results highlight the potential of using XGBoost in wearable systems, such as a smart bra, for breast tumor detection using ML. However, the high performance reported may be the result of overfitting, as the data were obtained from a single phantom, limiting the generalization of the findings. Elnaggar et al (2024) [104] proposed a wearable textile-based bra-tenna system for breast cancer detection, integrating UWB microwave sensing with ML algorithms. The classifiers evaluated included SVM, RF, gradient boosting machine, DT, AdaBoost, CatBoost, XGBoost, and LR. Two classification scenarios were studied: binary classification (healthy vs non-healthy) and multiclass classification (healthy, one tumor, and two tumors). The binary dataset comprised 1602 samples, and the multiclass dataset 2403 samples-both evenly distributed across classes. In the binary scenario, SVM achieved the best performance with 98% accuracy, 98% precision, 98% recall, 98% F1-score, and an AUC of 1.00. In the multiclass scenario, SVM again outperformed all other models, reaching 99% accuracy, 98.67% precision, 98.67% recall, and 98.67% F1-score. Patil et al (2024) [111] evaluated the performance of five ML algorithms-LR, SVM, MLP, KNN, and RF-for breast tumor detection. The dataset consisted of 804 feature vectors obtained from measurements on experimental breast phantoms. The class distribution was not disclosed. The dataset was split into training and test sets using a 70/30 ratio. Among the tested models, RF achieved the best performance, with 98.04% accuracy, 96.72% sensitivity, 96.67% specificity, 96.72% precision, 96.72% F1-score, and an AUC of 98.05%. Pelicano et al (2025) [112] investigated whether the complexity of breast phantoms affects the classification performance between benign and MTs. A total of eight numerical breast phantoms were used, from which 830 backscattered signals were collected, 460 benign and 370 malignant. The phantoms varied in the number of tissue types and whether those tissues were modeled as homogeneous or heterogeneous. An SVM classifier was employed for the analysis. The best performance was achieved using the phantom composed of homogeneous fat and fibroglandular tissue, with an accuracy of 98.8%, sensitivity of 97.3%, specificity of 100%, F1-score of 99.0%, and a MCC of 0.98. Vijayasarveswari et al (2025) [116] investigated various FE and selection techniques for breast cancer detection using ML algorithms, including DT, probabilistic neural network (PNN), and NB. Three FE methods were evaluated: K-means, reconstruction ICA, and statistical FE. For feature selection, the methods tested were Relief-F, neighborhood component analysis (NCA), and particle swarm optimization. The dataset consisted of 2000 samples obtained from an experimental breast phantom, evenly divided between tumorous and non-tumorous cases. Among the extraction methods, statistical features performed best, 18 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al achieving an average accuracy of 75.05% across the three classifiers. When combined with NCA as the feature selection method, the system achieved an improved average accuracy of 85.03%, making Statistical Features +NCA the best-performing combination. 3.1.2. Deep learning Woten et al (2007) [54] applied an ANN for breast cancer detection using a homogeneous breast numerical phantom containing a tumor. The dataset used in the study comprised 14 000 simulation cases: 4000 without tumor presence and 10 000 with tumor presence. The position of the tumor was varied across 100 random locations, and in both tumor and non-tumor (NT) cases, the electrical properties of the background were varied. The dataset was divided into training and test sets, each consisting of 7000 simulation cases. The ANN was trained and tested on these datasets under different noise conditions. Gaussian noise with a mean of zero and standard deviations of 0.001, 0.01, 0.1, 0.2, and 0.3 was added. The models were evaluated in two scenarios: one with binary classification (tumor present vs no tumor), where a single threshold of 0.5 was used to separate the two classes, and another with three-class classification (tumor present vs undecided vs no tumor), where a double threshold approach was applied, with cut-offs at 0.4 and 0.7: predictions below 0.4 were classified as ‘no tumor’, above 0.7 as ‘tumor present,’ and those between 0.4 and 0.7 as ‘undecided’. The best result for the binary classification, an accuracy of 97.59%, was achieved when the ANN was trained with noise of a standard deviation of 0.2. For the three-class classification, the best accuracy, 95.71%, was obtained with noise of standard deviation 0.1. Given the unbalanced dataset, additional performance metrics should be used to evaluate models performance, such as the F1-score or Matthew’s CORR (MCC). McGinley et al (2010) [64] proposed spiking neural network (SNN) as a tumor classification method. The SNN was employed along with a genetic algorithm (GA), with the parameters used for GA as detailed in [123]. In [64], the primary focus of the authors was the analysis of small tumors (up to 1 cm in radius). The tumor models used were based on the GRS model. A total of 368 tumor models were considered, comprising 184 of size 2.5 mm and 184 of size 7.5 mm. Among these tumors, 184 were of type 1 (malignant), 92 were of type 2 (macrolobulated benign), and 92 were of type 3 (smooth benign). The authors opted to employ LDA as a baseline for examining the performance and robustness of the SNN classifier. The dataset was randomly shuffled and divided into 75% for the training set and 25% for the test group. The algorithms were assessed using two types of classifier architectures: (i) a direct classifier that categorizes each tumor as either benign or malignant, and (ii) a two-stage classifier that initially categorizes each tumor as small or large, followed by categorization as benign or malignant. The results for the algorithms under the first classifier architecture showed an accuracy of 82.3% for LDA and 94.0% for SNN. For the two-stage classifier architecture, LDA demonstrated an accuracy of 91.2% for the size stage and 90.69% for the shape stage, while SNN exhibited 99.5% accuracy for the size stage and 98.71% for the shape stage. Also using SNN, in O’Halloran et al (2011) [67], the authors introduced a new approach using the classifier within a three-dimensional breast model with varying dielectric properties, using the result of applying the DWT to the RTS. Their dataset consisted of simulated signals of 160 tumor models obtained through the GRS method, including 80 malignant, 40 macrolobulated benign, and 40 smooth benign cases. The researchers employed GA techniques to optimize the parameters of the SNN and compared its performance with a LDA classifier. Results indicated that the SNN outperformed the LDA classifier and showed resilience to increasing levels of model heterogeneity, achieving 98% of accuracy, 100% of sensitivity and 95.76% of specificity. Yahya et al (2011) [68] implemented an ANN for breast cancer detection using a numerical breast phantom. The study considered tumors of three different sizes: 1 mm, 3 mm, and 5 mm. The dataset used to train and test the ANN comprised 1284 samples, with 1024 for training and 260 for testing. Each sample consisted of a feature vector obtained from the coefficients of a DWT. The paper did not disclose the class distribution or the number of samples corresponding to each tumor size. However, since the authors reported accuracy values for each tumor size separately, it is possible that 1284 samples were available for each tumor size. The highest classification accuracy was achieved for 5 mm tumors (100%), followed by 3 mm (76.56%) and 1 mm (65.52%). Jones et al (2013) [71] employed self-organizing maps (SOMs) to effectively classify tumor models. These tumor models encompassed three categories: macrolobulated benign, and two types of MTs with different degrees of spiculation. In this study, 90 tumor models were used, consisting of 30 BTs and 60 MTs, with the malignant group further divided into 30 tumors with 3 spicules and 30 tumors with 10 spicules. Two distinct datasets were analyzed: one with data from simulations featuring a tumor located in homogeneous breast tissue, and the other set where the tumor was located in heterogeneous breast tissue. Each dataset comprised 360 tumor signals. To evaluate the classifier, each dataset was randomly shuffled and divided into ten combinations of 276 training and 84 testing tumors. The classification process was iterated 10 times for each combination, and the average performance of the classifier was calculated. The results indicated an accuracy 19 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al of 99.5% for macrolobulated BTs, 90.54% for 3 spiculated MTs, and 88.28% for 10 spiculated MTs, resulting in an overall average accuracy of 92.77%. Caorsi et al (2017) [80] proposed a breast cancer detection approach using millimeter-wave signals and an ANN classifier. The authors employed four differentiated Gaussian pulses (DGP), each centered at a different frequency: 2 GHz, 6 GHz, 20 GHz, and 30 GHz. The dataset consisted of 200 samples, evenly split between Healthy (H) and Non-Healthy (NH) cases, and was divided equally into training and test sets (100 samples each). The best performance was achieved using data acquired with the 20 GHz DGP, where the ANN reached 92% accuracy, 96% sensitivity, and 88% specificity. Hirose et al (2019) [86] proposed a neural network-based method for breast cancer detection using the raw electric field frequency response as input. The dataset comprised 1400 samples generated from numerical breast phantoms, equally divided between healthy and tumorous cases (700 each). The neural network achieved an accuracy of 86.7% in classifying the presence or absence of breast cancer. Using the dataset from [121], Reimer et al (2022) [95] used a combined set of 1257 scans-1008 for training (augmented by a factor of 12) and 249 for testing-to evaluate CNN, DNN, and LR models. Among these, the CNN achieved the highest performance, with 75% accuracy, 82% sensitivity, and an AUC of 78%. Eashour et al (2024) [105] used the 1008-scan dataset and implemented a paired-breast CNN strategy with data augmentation through horizontal flips and translations. The paired-breast CNN outperformed the single-breast model, achieving an AUC of 85%, 75% sensitivity, and 79% specificity. To enhance model performance through data augmentation, Tariq et al (2025) [115] investigated the use of generative adversarial networks (GANs) to synthetically augment the first and second generations of the dataset. CNN models trained only on original data achieved 91.2% accuracy, 90.8% precision, 91.5% recall, and a 91.1% F1-score, whereas models trained solely on GAN-generated data showed slightly lower performance, with 89.7% accuracy, 89.1% precision, 90.2% recall, and an F1-score of 89.6%. However, combining original and synthetic data resulted in a marked improvement across all metrics, with 94.8% accuracy, 94.3% precision, 95.2% recall, and a 94.7% F1-score. The work by Tariq et al [115] highlights the potential of GAN-based data augmentation, which, when combined with real samples, led to significant gains in classification performance. The lowest performance was reported by Reimer et al [95], who trained models using a combination of the first and second generations. Halim et al (2023) [99] employed a PNN for breast cancer detection using a dataset of 10 000 samples obtained from an experimental breast phantom. The class distribution was not disclosed. The dataset was split into a training set (60%) and a test set (40%). The authors evaluated five feature normalization techniques in both the time and frequency domains: binary normalization, decimal scaling, linear scaling, min–max, and Z-score. The highest performance was achieved using frequency-domain features normalized with Z-score, reaching an accuracy of 98.67%. Lu et al (2023) [100] employed 1D CNN for breast cancer detection using backscattered signals generated from numerical breast phantoms. The dataset consisted of 5280 signals, evenly split between tumor-containing and healthy cases. The data were divided into a training set of 4224 signals and a test set of 1056 signals. The model achieved an accuracy of 98.20%, sensitivity of 96.52%, specificity of 100%, precision of 100%, and an F1-score of 98.23%. The authors also compared their results with previous studies by Liu et al [90] and Santorelli et al [74], showing that their 1D CNN approach outperformed these methods by more than 10 percentage points in terms of accuracy. Ghosh et al (2024) [108] employed a CNN based on the AlexNet architecture to classify breast tumors. The model was evaluated on two classification tasks: (1) distinguishing between tumorous and non-tumorous cases, and (2) differentiating between benign and MTs. The dataset consisted of 1150 rMWI images of numerical breast phantoms without tumors, 1152 with BTs, and 1152 with MTs. For the tumor vs NT classification, it achieved 97.22% accuracy, 97.01% precision, 97.10% recall, and an F1-score of 97.12%. For the benign vs malignant classification, the model achieved an accuracy of 98.59%, precision of 98.40%, recall of 98.52%, and F1-score of 98.55%. In Taghipour-Gorjikolaie et al (2024) [109], using data acquired using the Mammowave device, an autoencoder decoder PNN (AEDPNN) was used, but performance was weak due to data unbalance, same data than [110], with 59.7% accuracy, 58.75% sensitivity, and 59.88% specificity. Tables 8compile information on the equipment configurations used for signal acquisition in microwave breast imaging studies. The reviewed works employed a diverse range of antenna setups, from single-element configurations to dense arrays with up to 144 positions. Most systems use monostatic or bistatic arrangements, although quasi-multistatic configurations also appear in several studies; in some cases, multiple arrangements are implemented within the same system. While fixed platforms remain predominant, many recent studies incorporate moving antenna systems to enhance spatial diversity. Operating frequencies span from sub-GHz levels to over 20 GHz, with the majority of studies focused within the 1–10 GHz range. A clear trend toward increased antenna counts and broader frequency coverage is 20 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al Table 8. Summary table on the equipment used for signal acquisition (Target zone: breast). Label ‘T’ represents transmitters, while ‘R’ represents receivers. When these labels are absent, the respective study did not distinguish between transmitters and receivers. The (∗) indicates that the nomenclature used in this paper differs from that employed by the authors of the paper. ‘NA’ means not available. Reference No. of antenna positions Monostatic vs multistatic Moving vs fixed Frequencies Woten et al [54] 1 T 1 R Bistatic Fixed 6 GHz Davies et al [57] 8 Monostatic Fixed 6 GHz Chen et al [58] 2 T 5 R Multistatic Fixed 1, 4 GHz Chen et al [59] 2 T 5 R Multistatic Moving 1, 4 GHz Chen et al [60]1Monostatic Fixed 1 GHz Teo et al [61] 5, 7 Monostatic Fixed 6 GHz Conceiç˜ ao et al [62] 4 T 4 R Monostatic Fixed 6 GHz Kosmas et al [63] 7 Multistatic Fixed up to 10 GHz McGinley et al [64] 4 T 4 R Monostatic Fixed 6 GHz Conceiç˜ ao et al [65] 4 T 4 R Monostatic Fixed 6 GHz Conceiç˜ ao et al [66] 4 T 4 R Monostatic Fixed 6 GHz O’Halloran et al [67] 4 T 4 R Monostatic Fixed 6 GHz Yahya et al [68] 24 Multistatic Fixed 3.1–10.6 GHz Byrne et al [69] 4 T 4 R Multistatic Fixed 6 GHz Byrne et al [70] 4 T 4 R Monostatic Fixed 6 GHz Jones et al [71] 4 T 4 R Monostatic Fixed 6 GHz Santorelli et al [72] 16 Quasi-multistatic∗Fixed 2–4 GHz Conceiç˜ ao et al [73] 144 Monostatic Moving 1–6 GHz Santorelli et al [74] 16 Quasi-multistatic∗Fixed 2–4 GHz Reza et al [75] 1 T 1 R Bistatic Fixed 4.7 GHz Conceiç˜ ao et al [76] 144 Monostatic Moving 1–6 GHz Li et al [77] 16 Quasi-multistatic∗Fixed 2–4 GHz Sacristán et al [78] NA Multistatic Moving 2.3–6.5 GHz Gerazov et al [79] 4 T 4 R Monostatic Fixed 6 GHz Caorsi et al [80] NA Monostatic Fixed 20 GHz Oliveira et al [81] 12 Multistatic Fixed 6 GHz Aydin et al [82] 1 T 1 R Bistatic Fixed 3–18 GHz Aldhaeebi et al [83] 1 Monostatic Fixed 916 MHz Rana et al [84] 15 T 15 R Bistatic Moving 1–9 GHz Reimer et al [85] 1 T 12 R Bistatic Moving 2.3–6.5 GHz Hirose et al [86] 11 T 11 R Multistatic Fixed 0.207–4.14 GHz Conceiç˜ ao et al [55] 144 Monostatic Moving 1–6 GHz Rana et al [87] 15 T 80 R Multi-bistatic Moving 1–9 GHz Fasoula et al [88] 18 Multistatic Fixed 0.5–4 GHz Chen et al [89] 8 Multistatic Fixed 3 GHz Liu et al [90] 8 Quasi-Multistatic Fixed 6 GHz Sami et al [91] 72 T 72 R Bistatic Moving 2–4 GHz Patel et al [92] 72 T 72 R Monostatic/Bistatic Moving 1–8 GHz Martins et al [93] 1 Monostatic Fixed 2–6 GHz Janjic et al [94] 36 T 36 R Bistatic Moving 0.6–8 GHz Reimer et al 2022 [95] 72 Monostatic/Bistatic NA 1–8 GHz Janjic et al [96] 36 T 36 R Bistatic Moving 1–8 GHz Papini et al [97] 15 T 80 R Bistatic∗Moving 1–9 GHz Janjic et al [98] NA NA NA 0.5–8 GHz Halim et al [99] 2 Bistatic Fixed 4.3 GHz Lu et al [100] 12 Multistatic Fixed 6 GHz Ghavami et al [101] 15 T 80 R Bistatic Moving 1–9 GHz Rana et al [102] 5 T 80 R Bistatic Moving 1–9 GHz Abdelmawgood et al [103] 4 Quasi-multistatic Fixed 3.5–5 GHz Elnaggar et al [104] 4 Multistatic Fixed 2.5–12 GHz Eashour et al [105] 72 T 72 R Multistatic Moving 1–9 GHz Shadwell et al [106] 2 Multi-bistatic Moving 1–9 GHz (Continued.) 21 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al Table 8. (Continued.) Reference No. of antenna positions Monostatic vs multistatic Moving vs fixed Frequencies Dridi et al [107] NA Monostatic/Bistatic NA 1–8 GHz Ghosh et al [108] 1 Monostatic Fixed 2–10 GHz Taghipour-Gorjikolaie et al [109] 8 T 8 R Bistatic Moving 1–9 GHz Taghipour-Gorjikolaie et al [110] 10 T 80 R Bistatic Moving 1–9 GHz Patil et al [111] 1 Monostatic Fixed 2.76–11.14 GHz Pelicano et al [112]10 Monostatic Fixed 3 GHz Yurtseven et al [113] 36 T 36 R Bistatic Moving 0.6–8 GHz Reimer et al [114] NA NA NA NA Tariq et al [115] 72 T 72 R Monostatic/Bistatic Moving 1–8 GHz Vijayasarveswari et al [116] 2 Bistatic Fixed NA evident in more recent contributions. However, inconsistencies in reporting practices and terminology, along with the limited adoption of standardized measurement setups, continue to pose challenges for direct comparison across studies. The information regarding dataset sizes, input types, and model performance in the reviewed studies is summarized in table 9. These tables highlight a clear dominance of studies using feature vectors as model input, indicating widespread use of pre-processing techniques. Nevertheless, some studies use raw data directly from the acquisition devices, and in a few cases, images are used as input. The dataset sizes reported do not allow for conclusive analysis, as the number reported often refer to the number of raw measurements rather than the number of input samples actually used for training. Most studies employed balanced datasets, whereas the few that used unbalanced datasets often reported notably poorer performance. For instance, Taghipour-Gorjikolaie et al [109] used a highly unbalanced dataset-863 healthy and 161 unhealthy samples-and reported performance metrics around only 60%. Accuracy is the most commonly used evaluation metric; however, recent studies show a growing trend toward reporting multiple metrics, leading to a more robust and informative model evaluation. 3.2. Brain The applications of combining ML and MWI to classify brain diseases included the classification of strokes, classification of various stages of AD, and the classification of brain tumors. Table 10 presents a summary of papers on the application of ML algorithms for classifying brain images using MWI. 3.2.1. ML Salucci et al [124] pioneered the application of MWI for brain disease classification, specifically focusing on stroke detection using SVM. The authors trained SVM models on subsets of varying sizes (N=4 to 2500) drawn from a dataset of 2500 samples, which comprised 50% stroke cases, equally divided between IS and ICH, and 50% healthy cases. All data were acquired using experimental phantoms. Model performance was evaluated using an independent test set of 500 samples. Since the model achieved 100% accuracy on the test set for training sizes equal to or greater than N=5, only results up to that point were reported. While these results are promising, further validation using more complex phantoms and clinical data is necessary. In a related study, Guo et al [125] proposed a new approach for stroke localization and classification using MWI. Their pipeline consisted of three main steps: (i) tomographic reconstruction of the brain’s dielectric property profile using the BIM; (ii) FE via k-means clustering; and (iii) classification using an SVM model. Notably, the authors did not include healthy subjects in their analysis, assuming stroke diagnosis had already been made. They employed two MRI-derived numerical phantoms: Phantom A (256 ×256 ×128 voxels) and Phantom B (256 ×256 pixels). Classification was posed as a binary task distinguishing ICH from IS. Both training and testing datasets included 600 BIM-generated images (200 per SNR-level: 40 dB, 25 dB, 10 dB), with 100 images per class and SNR level. The SVM achieved 93.5% accuracy in low-noise conditions (40 dB) and over 81% accuracy in high-noise conditions (10 dB). When combining all SNR levels, the framework attained an 88% overall classification accuracy, demonstrating its robustness. Building upon this line of research, Zhu et al [56] introduced a novel method using graph degree mutual information (GDMI) to differentiate directly between ICH and IS from electromagnetic signals. They simulated 100 realistic brain models using segmented MRI data and an antenna array operating from 0.7 to 2 GHz. The time-domain signals were obtained via inverse Fast Fourier Transform, resulting in 16 ×16 time series per model. These were converted into graphs using the fast weighted horizontal visibility algorithm, with features extracted using GDMI and classified using an SVM. Using 50 ICH and 50 IS samples, the 22 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al Table 9. Summary of dataset characteristics and evaluation metrics. Reference Dataset Class balance Input sample Metric Woten et al [54] 14 000 (Signal) 4000/10 000 (H/NH) Signal Acc: 97.59% Davies et al [57] 90 (TM) 30/30/30 (S/micro/Sp) FV Acc:99.40% Chen et al [58] 60 (TM) 30/30 (Macro/micro) Signals ROC: NA Chen et al [59] 60 (TM) 30/30 (Macro/micro) Signals ROC: NA Chen et al [60] 60 (TM) 30/30 (Macro/micro) Signals ROC: NA Teo et al [61] 60 (TM) 30/30(S/Sp) Signals NA Conceiç˜ ao et al [62] 352 (TM) NA FV Acc: 85.99% Kosmas et al [63] NA (TM) NA (Macro/Micro) Signals NA McGinley et al [64] 368 (TM) 184/92/92 (Mal/MaB/SB) FV Acc: 98.71% Conceiç˜ ao et al [65] 352 (TM) 176/176 (Ben/Mal) FV Acc: 87.05% Conceiç˜ ao et al [66] 480 (TM) NA FV Acc: 92.71%, 73.96% O’Halloran et al [67] 160 (TM) 80/40/40 (Mal/MaB/SB) FV Acc, Sen, Spe: 98%, 100%, 95.76% Yahya et al [68] 1284 (FV) NA FV Acc (5 mm, 3 mm, 1 mm): 100%, 76.56%, 65.52% Byrne et al [69] 1728 (FV) NA FV Acc: 83.66% Byrne et al [70] 2880 (Signals) 1440/1440 (H/NH) FV Acc: 94.95% Jones et al [71] 90 (TM) NA FV 92.77% Santorelli et al [72] 230 (Breast scans) 115/115 (H/NH) FV Acc: 73.64% Conceiç˜ ao et al [73] 26 (TM) 13/13 (Ben/Mal) FV Acc: 89.34% Santorelli et al [74] 320 50%/50% (H/NH) FV Acc, FP, FN: 77,53%, 8.33%, 37.41% Reza et al [75] 300 36/264 (H/NH) FV Acc, Sen, Spe: 100%, 100%, 100% Conceiç˜ ao et al [76] 26 (TM) 13/13 (Ben/Mal) FV Acc: 90.95% Li et al [77] 290 (Phantoms) 150/140 (H/NH) FV AGE: 0.129 Sacristán et al [78] 1534/2395 (Breast models) 50%/50% (H/NH) FV Acc: 94%, 62% Gerazov et al [79] 240 (TM) 80/160 (Ben/Mal) FV Acc: 92.81% Caorsi et al [80] 200 (NA) 100/100 (H/NH) FV Acc, Sen, Spe: 92%, 96%, 88% Oliveira et al [81] 84 240 (FV) NA FV ROC: NA Aydin et al [82] 400 (signals) NA FV Acc: 94% Aldhaeebi et al [83] 90 (Breast model) 45/45 (H/NH) FV Acc, Rec, Spe, Prec: 65%, 55%, 75%, 68.8% Rana et al [84] 23 (Breasts) 12/11 (H/NH) NA Acc, Sen, Spe: 98.9%, 97,7%, 99.7% NA Reimer et al [85] 2000/2001 (Phantoms) 50%/50% (H/Mal) FV Sen, Spe, AUC: 83%, 91%; 94% Hirose et al [86] 1400 (NA) 700/700 (H/NH) Scattered field Acc: 85% Conceiç˜ ao et al [55] 26 (TM) 13/13 (Ben/Mal) FV Acc: 96.2%, 92.3% Rana et al [87] 61 (Breast exams) 25/36 (H/NH) FV Acc, Sen, Spe: 91%, 84.40%, 97.20% Fasoula et al [88] 24 (Patients) 11/8/5 (BPBC/UC/BPBBL) Solidity, Correlation, Busyness CL: 11.50% Chen et al [89] 600 (Signal) NA Signal Acc, F1: 72.5%, 72.8% Liu et al [90] 11 232 (Signals) 5616/5616 (H/NH) FV Acc: 84.8% Sami et al [91] 1008 (Matrix S21) NA Signal Acc, Sen, Spe: 99.7%, 99.2%, 99.9% Patel et al [92] 249 (Breast Scans) NA FV Acc, Sen, Spe, AUC: 94%, 100%, 89.2%, 94% Martins et al [93] 480 (Breast scans) 240/240 (H/NH) FV Acc: 80% Janjic et al [94] 54 (Breast exams) 29/25 (Ben/Mal) Sii Matrix Acc, Sen, Spe: 81%, 80%, 83% Reimer et al [95] 1257 (Breast scans) NA Matrix S11 Acc, Sen, Spe, AUC: 75%, 82%, 70%, 78% Janjic et al [96] 113 (Breast exams) 70/43 (Ben/Mal) Sii Matrix Acc, Sen, Spe: 78%, 79%, 77% (Continued.) 23 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al Table 9. (Continued.) Reference Dataset Class balance Input sample Metric Papini et al [97] 697 (Breast exams) 574/123 (H/NH) Signals/FV Acc, Sen, Spe: 88%, 86%, 89% 83%, 77%, 85% Janjic et al [98] 558 (Breast scans) 390/168 (H/NH) Signal Acc, Sen, Spe: 93%, 95%, 92% Halim et al [99] 10 000 (FV) NA FV Acc: 98.67% Lu et al [100] 5280 (Signal) NA Signal Acc, Sen, Spe, Prec, F1: 98.20%, 96.52%, 100%, 100%, 98.23% Ghavami et al [101] 697 (Breast scans) 321/376 (Ben/Mal) FV Acc, Sen, Spe: 90%, 90%, 92% Rana et al [102] 61 (Breasts) 26/35 (Ben/Mal) FV Acc, Sen, Spe: 91%, 84%, 95% Abdelmawgood et al [103] 1001 (Signal) 50%/50% (H/NH) FV Acc: 98.5% Elnaggar et al [104] 1602/2403 (FV) 801/801; 801/801/801 (H/NH; H/1 T/2 T) FV Acc, Rec, Prec, F1: 98%, 98%, 98%, 98%; 99%, 98.67%, 98.67%, 98.67% Eashour et al [105] 4030 (Images) NA Images Sen, Spe, AUC: 75%, 79%, 85% Shadwell et al [106] 352 120/170/62 (H/Ben/Mal) FV Acc: 80% Dridi et al [107] 249 (Breast scans) NA FV Acc, Sen, Spe: 98%, 96.15%, 100% Ghosh et al [108] 3454/2304 (Images) 1150/2304 (H, NT); 1152/1152 (Ben, Mal) Images Acc, Rec, Prec, F1: 97.22%, 97.10%, 97.01%, 97.12%; 98.59%, 98.52%, 98.40%, 98.55% Taghipour-Gorjikolaie et al [109] 1026 (Breast scans) 863/161 (H/NH) FV Acc, Sen, Spe: 59.70%, 58.75%, 59.88% Taghipour-Gorjikolaie et al [110] 1026 (Breast scans) 863/161 (H/NH) FV Acc, Sen, Spe: 64.49%, 61.25%, 65.08% Patil et al [111] 804 (FV) NA FV Acc, Sen, Spe, Prec, F1, AUC: 98.04%, 96.72%, 96.67%, 96.72%, 96.72%, 98.05% Pelicano et al [112] 830 (Observations) 460/370 (Ben/Mal) FV Acc, Sen, Spe, F1, MCC: 98.8%, 97.3%, 100%, 99.0%, 0.98 Yurtseven et al [113] 526 (Patients) 286/240 (Ben+H/Mal) Matrices S11 and S12 Acc, Sen, Spe: 91%, 91%, 92% Reimer et al [114] 200 (Breast Scans) NA FV AUC: NA Tariq et al [115] 1257 (Breast Scans) NA Spectograms Acc, Rec, Prec, F1: 94.8%, 95.2%, 94.3%, 94.7% Vijayasarveswari et al [116] 2000 (data samples) 1000/1000 (H/NH) FV Acc: 85.03% 1 Tumor (1 T), 2 Tumor (2 T), accuracy (Acc), average generalization error (AGE), benign (Ben), biopsy-proven benign breast lesions (BPBBL), biopsy-proven breast cancer (BPBC), false negative rate (FN), false positive rate (FP), feature vector (FV), healthy (H), macrolobulated (Macro), macrolobulated Benign (MaB), Malignant (Mal), Matthew’s correlation coefficient (MCC), microlobulated (Micro), not available (NA), Not Healthy (NH), precision (Prec), recall (Rec), receiver operating characteristic (ROC), sensitivity (Sen), Smooth (S), smooth benign (SB), specificity (Spe), spiculated (Sp), tumor models (TM), unaspirated cysts (UC). method achieved 91% sensitivity, 98% specificity, and 94% accuracy in noiseless conditions. Under noise levels of 40 dB, 25 dB, and 10 dB, accuracies dropped to 93%, 88%, and 77%, respectively, with an overall accuracy of 89%. These results align with those of Guo et al [125], confirming that noise degrades classification performance, although Guo et al’s method maintained a slight accuracy advantage. Complementing these earlier efforts, Pokorny et al [130] have conducted a series of studies applying SVM to classify brain strokes using MWI data derived from numerical head models. The authors investigated multi-class classification using SVMs with data from numerical head models. They introduced a new class-no stroke (noStroke)-resulting in a three-class problem: IS, ICH, and noStroke. They used head models from the IT’IS Foundation, to create three datasets: (1) 1000 simulations per class with predefined stroke sizes and positions; (2) 200 simulations per class with random sizes and fixed positions; and (3) 200 simulations per class with both random sizes and positions. Four hypotheses regarding generalization from 24 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al Table 10. Classification using machine learning models (Target zone: brain). Reference Algorithm Type of Data Salucci et al [124] SVM Numerical Guo et al [125] SVM Numerical Zhu et al [56] SVM Numerical Saied et al [126] LR, LDA, KNN, DT, NB, SVM Numerical Hossain et al [127] YOLOv5 Experimental Lalitha et al [128] SVM, KNN, Bagging, RF, J48, NB, DT Numerical Hossain et al [129] MBINet Experimental Pokorny et al [130] SVM Numerical Pokorny et al [131] SVM Numerical Ullah et al [132] KNN RF DT Numerical Hossain et al [133] MSegNet, BINet Experimental Cardinali et al [134] MLP Experimental Taha et al [135] BrainVisionNet Experimental Hasan et al [136] SVM, KNN, RF, ANN Numerical Hasan et al [137] RF, SVM, XGBoost Numerical Hashir et al [138] TinyML Numerical Cardinali et al [139] SVM Experimental Hossain et al [140] FT-FEDTL Experimental Hasan et al [141] CNN Clinical Pokorny et al [142] SVM Numerical Lalitha et al [143] YOLOv5l Numerical Farhatullah et al [144] CNN Numerical Sudhakaran et al [145] MobileNet v2 Experimental Sudhakaran et al [146] Hybrid Inception-CNN Experimental Yuan et al [147] ANN Clinical Artificial neural network (ANN), BrainImageNet (BINet), convolutional neural network (CNN), decision tree (DT), fine-tuned feature-extracted deep transfer learning (FT-FEDTL), k-nearest neighbors (KNN), linear discriminant analysis (LDA), logistic regression (LR), microwave brain image network (MBINet), MicrowaveSegNet (MSegNet), multi-layer percepton (MLP), Naïve Bayes (NB), random forest (RF), support vector machines (SVM). small stroke data were tested. When trained on small stroke data, SVMs achieved 95.7% accuracy on similar stroke sizes but only 65.5% on larger strokes. Surprisingly, using multi-frequency data led to worse results (33.3% accuracy with 25 frequencies) than single-frequency input (94.6%), but PCA-based dimensionality reduction improved performance across all cases (up to 96.9%). For strokes with randomized size and position, accuracy dropped to a maximum of 70.5%, indicating that training data diversity and input dimensionality significantly affect model robustness. In another study, Pokorny et al [131] used 2D datasets for multi-class classification and compared SVM with other classifiers, including LR, discriminant analysis, KNN, NB, and DT. They employed Bayesian optimization for hyperparameter tuning and found that increasing the separation between adjacent antennas improved performance-accuracy rose from 66.2% to 68.7% and Cohen’s kappa from 0.24 to 0.29. Though the overall accuracy was lower than in their previous 3D study, the work provided insight into how antenna configuration impacts classification outcomes. In 2024, Pokorny et al [142] presented a systematic evaluation of SVM performance in both binary and three-class classification. Two datasets were used: 3D_2 for training (600 samples across noStroke, IS, and ICH), and 3D_3 for testing (300 samples). In the binary case (noStroke vs IS), SVM achieved 86.7% accuracy. In the three-class model, it reached 86.3%. These findings underscore the effectiveness of SVM for MWI-based stroke classification, particularly when class balance and training data quality are well controlled. The first study of AD using MWI data was presented by Saied et al (2021) [126]. They used five ML algorithms, namely LR, KNN, DT, NB, and SVM, with the aim of classifying various stages of AD using data obtained from numerical model simulations. Additionally, LDA was used in their study. These models were developed in the CST Microwave Studio Suite, with measurements acquired from previous studies [148, 149]. Each stage of AD (normal, Mild AD, Moderate AD, and Severe AD) was represented by changes in the dielectric properties of specific regions and tissues. Simulations were conducted using six antennas, and the signals from each antenna were aggregated into a single dataset for each simulation case. This dataset was then divided into a training set, comprising approximately 78% of the complete simulation cases, and a validation set, which accounted for the remaining 22%. The results of the study indicated that LR achieved 25 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al [22] Ambrosanio M et al 2020 Artificial neural networks for quantitative microwave breast imaging Bioimaging 2204–8 [23] Mojabi P, Hughson M, Khoshdel V, Jeffrey I and LoVetri J 2021 CNN for compressibility to permittivity mapping for combined ultrasound-microwave breast imaging IEEE J. Multiscale Multiphys. Comput. Tech. 662–72 [24] Ambrosanio M, Franceschini S, Pascazio V and Baselice F 2022 An end-to-end deep learning approach for quantitative microwave breast imaging in real-time applications Bioengineering 9651 [25] Ambrosanio M, Autorino M M, Franceschini S, Baselice F and Pascazio V 2022 Microwave breast imaging via deep learning 2022 IEEE 19th Int. Symp. Biomedical Imaging (ISBI) (IEEE) pp 1–4 [26] Costanzo S, Flores A and Buonanno G 2022 Machine learning approach to quadratic programming-based microwave imaging for breast cancer detection Sensors 22 4122 [27] Costanzo S, Flores A and Buonanno G 2022 Machine learning methods for microwave imaging in cancer detection 2022 IEEE Int. Conf. on Dependable, Autonomic and Secure Computing, Int. Conf. on Pervasive Intelligence and Computing, Int. Conf. on Cloud and Big Data Computing, Int. Conf. on Cyber Science and Technology Congress (DASC/PiCom/CBDCom/CyberSciTech) (IEEE) pp 1–5 [28] No¨ el V, Qin Y, Rodet T and Lesselier D 2022 Breast imaging by cascaded CNN from joint microwave and ultrasonic data 2022 30th European Signal Processing Conf. (EUSIPCO) (IEEE) pp 917–21 [29] Qin Y, Ran P, Rodet T and Lesselier D 2022 Breast imaging by convolutional neural networks from joint microwave and ultrasonic data IEEE Trans. Antennas Propag. 70 6265–76 [30] Fontaine G and Pistorius S 2023 Machine learning based reconstruction of point-like scatterers in a portable microwave detection device 17th European Conf. Antennas and Propagation (EuCAP) (IEEE) pp 1–5 [31] Bicer M B 2023 Radar-based microwave breast imaging using neurocomputational models Diagnostics 13 930 [32] Costanzo S and Flores A 2023 CVNN-based microwave imaging approach 2023 IEEE Conf. Antenna Measurements and Applications (CAMA) (IEEE) pp 728–31 [33] Khoshdel V, Mojabi P and LoVetri J 2023 A multi-branch deep learning architecture for microwave-ultrasound breast imaging 2023 35th General Assembly and Scientific Symp. Int. Union of Radio Science (URSI GASS) (IEEE) pp 1–4 [34] Borghouts M, Ambrosanio M, Franceschini S, Autorino M M, Pascazio V and Baselice F 2023 Microwave breast sensing via deep learning for tumor spatial localization by probability maps Bioengineering 10 1153 [35] Franceschini S, Autorino M M, Ambrosanio M, Pascazio V and Baselice F 2023 A deep learning approach for diagnosis support in breast cancer microwave tomography Diagnostics 13 1693 [36] Ambrosanio M, Borghouts M, Franceschini S, Autorino M M, Pascazio V and Baselice F 2024 Enhanced deep-learning-based microwave sensing technology for breast cancer localization 2024 IEEE Int. Conf. E-Health Networking, Application & Services (HealthCom) (IEEE) pp 1–5 [37] Conceiç˜ ao R C, Byrne D, Noble J A and Craddock I 2016 Initial study for the investigation of breast tumour response with classification algorithms using a microwave radar prototype 2016 10th European Conf. Antennas and Propagation (EuCAP) pp 1–2 [38] Burfeindt M J, Colgan T J, Mays R O, Shea J D, Behdad N, Van Veen B D and Hagness S C 2012 MRI-derived 3-D-printed breast phantom for microwave breast imaging validation IEEE Antennas Wirel. Propag. Lett. 11 1610–3 [39] Xiao L-Y, Hong R, Zhao L-Y, Hu H-J and Liu Q H 2022 A hybrid neural network electromagnetic inversion scheme (HNNEMIS) for super-resolution 3-D microwave human brain imaging IEEE Trans. Antennas Propag. 70 6277–86 [40] Cheng Y, Xiao L-Y, Zhao L-Y, Hong R and Liu Q H 2022 A 3-D full convolution electromagnetic reconstruction neural network (3-D FCERNN) for fast super-resolution electromagnetic inversion of human brain Diagnostics 12 2786 [41] Zhao L-Y, Xiao L-Y, Cheng Y, Hong R and Liu Q H 2022 Machine-learning-based inversion scheme for super-resolution three-dimensional microwave human brain imaging IEEE Antennas Wirel Propag. Lett. 21 2437–41 [42] Mousavi S S S and Majedi M S 2023 High quality brain image reconstruction based on DBIM and U-net 2023 30th National and 8th Int. Iranian Conf. Biomedical Engineering (ICBME) (IEEE) pp 171–5 [43] Costanzo S, Flores A and Buonanno G 2023 Microwave imaging for brain cancer detection: enhanced accuracy with machine learning approach 2023 IEEE Int. Conf. Metrology for EXtended Reality, Artificial Intelligence and Neural Engineering (MetroXRAINE) (IEEE) pp 519–24 [44] Costanzo S, Flores A and Buonanno G 2023 Machine learning approach to enhanced resolution of inverse scattering for cancer detection 2023 Photonics & Electromagnetics Research Symp. (PIERS) (IEEE) pp 1692–7 [45] Flores A, Buonanno G and Costanzo S 2023 Machine learning approach to microwave imaging for cancer detection 2023 17th European Conf. Antennas and Propagation (EuCAP) (IEEE) pp 1–4 [46] Xue F, Guo L, Bialkowski A and Abbosh A 2024 Evaluation of fully convolutional networks for dielectric profile reconstruction in medical microwave imaging 2024 IEEE Int. Symp. Antennas and Propagation and INC/USNC-URSI Radio Science Meeting (AP-S/INC-USNC-URSI) (IEEE) pp 2383–4 [47] Zuo T, Jiang L, Cheng Y, Yu X, Tao X, Zhang Y and Cao R 2024 Deep learning-based electric field enhancement imaging method for brain stroke Sensors 24 6634 [48] Zuo T, Tao X, Jiang L, Chen Y, Zhang Y and Cao R 2024 Microwave stroke imaging system using learning electric field enhancement 2024 IEEE 12th Asia-Pacific Conf. Antennas and Propagation (APCAP) (IEEE) pp 1–2 [49] Costanzo S and Flores A 2024 CVNN approach for microwave imaging applications in brain cancer: preliminary results 2024 18th European Conf. Antennas and Propagation (EuCAP) (IEEE) pp 1–3 [50] Xue F, Guo L and Abbosh A 2025 Enhanced learning in microwave medical imaging using boundary-overlap-size loss 2025 6th Australian Microwave Symp. (AMS) (IEEE) pp 1–2 [51] Movafagh M, Ghavami N, Taghipour-Gorjikolaie M, Tiberi G, Cosottini M, Dudley S and Ghavami M 2025 Stroke classification via microwave imaging using Huygens’ principle assisted by deep learning 2025 19th European Conf. Antennas and Propagation (EuCAP) (IEEE) pp 1–3 [52] Mousavi S S S and Majedi M S 2025 Reconstruction and classification of brain strokes using deep learning-based microwave imaging IEEE Access vol 3, pp 27024–36 [53] Zubal I G, Harrell C R, Smith E O, Rattner Z and Gindi G 1994 Computerized three-dimensional segmented human anatomy Med. Phys. 21 299–302 [54] Woten D A, Lusth J and El-Shenawee M 2007 Interpreting artificial neural networks for microwave detection of breast cancer IEEE Mcrowave Wirel. Compon. Lett. 17 825–7 [55] Conceiç˜ ao R C, Medeiros H, Godinho D M, O’Halloran M, Rodriguez-Herrera D, Flores-Tapia D and Pistorius S 2020 Classification of breast tumor models with a prototype microwave imaging system Med. Phys. 47 1860–70 [56] Zhu G, Bialkowski A, Guo L, Mohammed B and Abbosh A 2021 Stroke classification in simulated electromagnetic imaging using graph approaches IEEE J. Electromagn. RF Microw. Med. Biol. 546–53 32 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al [57] Davis S K, Van Veen B D, Hagness S C and Kelcz F 2008 Breast tumor characterization based on ultrawideband microwave backscatter IEEE Trans. Biomed. Eng. 55 237–46 [58] Chen Y, Craddock I J, Kosmas P, Ghavami M and Rapajic P 2009 Application of the MIMO radar technique for lesion classification in UWB breast cancer detection European Signal Processing Conf. (IEEE) pp 759–63 [59] Chen Y, Craddock I J, Kosmas P, Ghavami M and Rapajic P 2010 Multiple-input multiple-output radar for lesion classification in ultrawideband breast imaging IEEE J. Sel. Top. Signal Process. 4187–201 [60] Chen Y, Craddock I J and Kosmas P 2010 Feasibility study of lesion classification via contrast-agent-aided UWB breast imaging IEEE Trans. Biomed. Eng. 57 1003–7 [61] Teo J, Chen Y, Soh C B, Gunawan E, Low K S, Putti T C and Wang S C 2010 Breast lesion classification using ultrawideband early time breast lesion response IEEE Trans. Antennas Propag. 58 2604–13 [62] Conceiç˜ ao R C, O’Halloran M, Glavin M and Jones E 2010 Support vector machines for the classification of early-stage breast cancer based on radar target signatures Prog. Electromagn. Res. B 23 311–27 [63] Kosmas P, Laranjeira S, Dixon J H, Li X and Chen Y 2010 Time reversal microwave breast imaging for contrast-enhanced tumor classification 2010 Annual Int. Conf. IEEE Engineering in Medicine and Biology Society, EMBC’10 (IEEE) pp 708–11 [64] McGinley B, O’Halloran M, Conceicao R C, Morgan F, Glavin M and Jones E 2010 Spiking neural networks for breast cancer classification using radar target signatures Prog. Electromagn. Res. C 17 79–94 [65] Conceiç˜ ao R C, O’Halloran M, Glavin M and Jones E 2011 Evaluation of features and classifiers for classification of early-stage breast cancer J. Electromagn. Waves Appl. 25 1–14 [66] Conceiç˜ ao R C, O’Halloran M, Glavin M and Jones E 2011 Effects of dielectric heterogeneity in the performance of breast tumour classifiers Prog. Electromagn. Res. M 17 73–86 [67] O’Halloran M, Cawley S, McGinley B, Conceicao R, Morgan F, Jones E and Glavin M 2011 Evolving spiking neural network topolo-gies for breast cancer classification in a dielectrically heterogeneous breast Prog. Electromagn. Res. Lett. 25 153–62 [68] Yahya A F, Abbosh Y M and Abbosh A 2011 Microwave imaging method employing wavelet transform and neural networks for breast cancer detection Asia-Pacific Microwave Conf. 2011 (IEEE) pp 1418–21 [69] Byrne D, O’Halloran M, Jones E and Glavin M 2011 Support vector machine-based ultrawideband breast cancer detection system J. Electromagn. Waves Appl. 25 1807–16 [70] Byrne D, O’Halloran M, Glavin M and Jones E 2011 Breast cancer detection based on differential ultrawideband microwave radar Prog. Electromagn. Res. M 20 231–42 [71] Jones M, Byrne D, McGinley B, Morgan F, Glavin M, Jones E, O’Halloran M and Conceiç˜ ao R C 2013 Classification and monitoring of early stage breast cancer using ultra wide band radar The Eighth Int. Conf. on Systems (ICONS) pp 46–51 [72] Santorelli A, Porter E, Kirshin E, Liu Y J and Popovi´ c M 2014 Investigation of classifiers for tumor detection with an experimental time-domain breast screening system Prog. Electromagn. Res. 144 45–57 [73] Conceiç˜ ao R C, Medeiros H, O’Halloran M, Rodriguez-Herrera D, Flores-Tapia D and Pistorius S 2013 Initial classification of breast tumour phantoms using a UWB radar prototype 2013 Int. Conf. Electromagnetics in Advanced Applications (ICEAA) pp 720–3 [74] Santorelli A, Li Y, Porter E, Popovi´ c M and Coates M 2014 Investigation of classification algorithms for a prototype microwave breast cancer monitor The 8th European Conf. Antennas and Propagation (EuCAP 2014) (IEEE) pp 320–4 [75] Reza K J, Khatun S, Jamlos M F, Fakir M M and Mostafa S 2014 Performance evaluation of diversified SVM kernel functions for breast tumor early prognosis ARPN J. Eng. Appl. Sci. 9329–35 [76] Conceiç˜ ao R C, Medeiros H, O’Halloran M, Rodriguez-Herrera D, Flores-Tapia D and Pistorius S 2014 SVM-based classification of breast tumour phantoms using a UWB radar prototype system 2014 31th URSI General Assembly and Scientific Symp. (URSI GASS) (IEEE) pp 1–4 [77] Li Y, Santorelli A, Laforest O and Coates M 2015 Cost-sensitive ensemble classifiers for microwave breast cancer detection ICASSP, IEEE Int. Conf. Acoustics, Speech and Signal Processing - Proc. (IEEE) pp 952–6 [78] Sacristán J, Oliveira B L and Pistorius S 2016 Classification of electromagnetic signals obtained from microwave scattering over healthy and tumorous breast models 2016 IEEE Canadian Conf. Electrical and Computer Engineering (CCECE) (IEEE) pp 1–5 [79] Gerazov B and Conceicao R C 2017 Deep learning for tumour classification in homogeneous breast tissue in medical microwave imaging IEEE EUROCON 2017-17th Int. Conf. Smart Technologies (IEEE) pp 564–9 [80] Caorsi S and Lenzi C 2017 Can a MM-wave ultra-wideband ANN-based radar data processing approach be used for breast cancer detection? 2017 Int. Conf. Electromagnetics in Advanced Applications (ICEAA) (IEEE) pp 1236–9 [81] Oliveira B L, Godinho D, O’Halloran M, Glavin M, Jones E and Conceicao R C 2018 Diagnosing breast cancer with microwave technology: remaining challenges and potential solutions with machine learning Diagnostics 81–21 [82] Avs¸ar Aydin E and Keles¸ M K 2019 Uwb rectangular microstrip patch antenna design in matching liquid and evaluating the classification accuracy in data mining using random forest algorithm for breast cancer detection with microwave J. Electr. Eng. Technol. 14 2127–36 [83] Aldhaeebi M, Bamatraf S, Ramahi O and Binajjaj S A 2019 Breast tumor diagnosis using machine learning with microwave probes 2019 First Int. Conf. Intelligent Computing and Engineering (ICOICE) (IEEE) pp 1–4 [84] Rana S P, Dey M, Tiberi G, Sani L, Vispa A, Raspa G, Duranti M, Ghavami M and Dudley S 2019 Machine learning approaches for automated lesion detection in microwave breast imaging clinical data Sci. Rep. 910510 [85] Reimer T, Sacristan J and Pistorius S 2019 Improving the diagnostic capability of microwave radar imaging systems using machine learning 2019 13th European Conf. Antennas and Propagation (EuCAP) (IEEE) pp 1–5 [86] Hirose U and Kidera S 2019 Breast tumor characterization with raw data based machine learning for microwave ultra-wideband mammography 2019 Int. Symp. Antennas and Propagation (ISAP) (IEEE) pp 1–3 [87] Rana S P, Dey M, Loretoni R, Duranti M, Sani L, Vispa A, Ghavami M, Dudley S and Tiberi G 2021 Radial basis function for breast lesion detection from mammowave clinical data Diagnostics 11 1930 [88] Fasoula A, Duchesne L, Cano J D G, Moloney B M, Elwahab S M and Kerin M J 2021 Automated breast lesion detection and characterization with the wavelia microwave breast imaging system: methodological proof-of-concept on first-in-human patient data Appl. Sci. 11 9998 [89] Chen A, Gu Y and Zhang S 2021 SVM-based microwave breast tumour classification 2021 Int. Conf. Public Health and Data Science (ICPHDS) (IEEE) pp 174–7 [90] Liu G, Xiao X, Song H and Kikkawa T 2021 Precise detection of early breast tumor using a novel EEMD-based feature extraction approach by UWB microwave Med. Biol. Eng. Comput. 59 721–31 33 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al [91] Sami H, Sagheer M, Riaz K, Mehmood M Q and Zubair M 2021 Machine learning-based approaches for breast cancer detection in microwave imaging 2021 IEEE USNC-URSI Radio Science Meeting (Joint With AP-S Symp.) (IEEE) pp 72–73 [92] Patel P and Raina A 2021 Comparison of machine learning algorithms for tumor detection in breast microwave imaging 2021 11th Int. Conf. Cloud Computing, Data Science & Engineering (Confluence) (IEEE) pp 882–6 [93] Martins R A, Felício J M, Costa J R and Fernandes C A 2021 Comparison of slot-based and Vivaldi antennas for breast tumor detection using machine learning and microwave imaging algorithms 2021 15th European Conf. Antennas and Propagation (EuCAP) (IEEE) pp 1–5 [94] Janjic A, Akduman I, Cayoren M, Bugdayci O and Aribal M E 2022 Gradient-boosting algorithm for microwave breast lesion classification-SAFE clinical investigation Diagnostics 12 3151 [95] Reimer T and Pistorius S 2021 The diagnostic performance of machine learning in breast microwave sensing on an experimental dataset IEEE J. Electromagn. RF Microw. Med. Biol. 6139–45 [96] Janjic A, Akduman I, Cayoren M, Bugdayci O and Aribal M E 2023 Microwave breast lesion classification - results from clinical investigation of the SAFE microwave breast cancer system Acad. Radiol. 30 S1–S8 [97] Papini L et al 2023 Breast cancer detection using machine learning approaches on microwave-based data 17th European Conf. Antennas and Propagation (EuCAP) pp 1–5 [98] Janjic A, Akduman I, Cayoren M, Bugdayci O and Aribal M E 2023 Support vector machine algorithm for clinical microwave breast cancer screening and early detection 2023 IEEE Conf. Antenna Measurements and Applications (CAMA) (IEEE) pp 344–6 [99] Halim A A A, Veeraperumal V, Andrew A M, Yasin M N M, Ahmad M Z Z, Hossain K, Bari B S and Kamal F 2023 UWB-based early breast cancer existence prediction using artificial intelligence for large data set J. Adv. Res. Appl. Sci. Eng. Technol. 29 81–90 [100] Lu M, Xiao X, Liu G, Lu H, Pang Y and Kikkawa T 2022 Breast tumor detection by 1D-convolutional neural network based on ultra-wide-band microwave technology Meas. Sci. Technol. 34 025702 [101] Ghavami N et al 2023 Mammowave breast imaging device: prospective clinical trial results and ai enhancement 2023 IEEE Conf. Antenna Measurements and Applications (CAMA) (IEEE) pp 341–3 [102] Rana S P, Dey M, Loretoni R, Duranti M, Ghavami M, Dudley S and Tiberi G 2023 Radiation-free microwave technology for breast lesion detection using supervised machine learning model Tomography 9105–29 [103] Abdelmawgood Z, Abdalawy S, Mohamed H, Eldamak A, Fahmy O and Elsheakh D 2023 Modeling of textile antenna sensors for breast cancer detection using artificial intelligence techniques 2023 Int. Conf. Modeling, Simulation & Intelligent Computing (MoSICom) (IEEE) pp 383–6 [104] Elnaggar A H, El-Hameed A S A, Yakout M A and Areed N F 2024 Machine learning for breast cancer detection with dual-port textile uwb mimo bra-tenna system Information 15 467 [105] Eashour F and Pistorius S 2024 Neural network based microwave tumour detection using breast pairs 2024 18th European Conf. Antennas and Propagation (EuCAP) (IEEE) pp 1–5 [106] Shadwell H, Nnadi S N, Aliyu A, Ghavami N, Ghavami M, Tiberi G and Sohani B 2024 Machine learning techniques for autonomous lesion detection in microwave breast imaging clinical data 2024 18th Int. Symp. Medical Information and Communication Technology (ISMICT) (IEEE) pp 95–98 [107] Dridi M and Gharsalli L 2024 Supervised machine learning for breast cancer detection using microwave imaging in the frequency domain 2024 18th European Conf. Antennas and Propagation (EuCAP) (IEEE) pp 1–4 [108] Ghosh M and Basu B 2024 Detection of malignant breast tissue using SAR observation with microwave imaging and convolutional neural network Comput. Electr. Eng. 119 109620 [109] Taghipour-Gorjikolaie M, Ghavami N, Tiberi G, Badia M, Papini L, Fracassini A, Bigotti A, Palomba G and Ghavami M 2024 Microwave-based breast cancer detection by using auto encoder-decoder probabilistic neural network 2024 18th Int. Symp. Medical Information and Communication Technology (ISMICT) (IEEE) pp 47–52 [110] Taghipour-Gorjikolaie M, Khalesi B, Ghavami N, Tiberi G, Badia M, Papini L, Fracassini A, Bigotti A, Palomba G and Ghavami M 2024 Frequency selection to improve the performance of microwave breast cancer detecting support vector model by using genetic algorithm 2024 IEEE Int. Symp. Medical Measurements and Applications (MeMeA) (IEEE) pp 1–6 [111] Patil S and Naik A 2024 UWB resonator-based supervised learning for breast tumor diagnosis Prog. Electromagn. Res. C 140 93–104 [112] Pelicano A C, Ara´ ujo N A, Godinho D M and Conceiç˜ ao R C 2025 A preliminary study on the impact of model complexity in classification in breast microwave imaging 2025 19th European Conf. Antennas and Propagation (EuCAP) (IEEE) pp 1–5 [113] Yurtseven A, Janjic A, Cayoren M, Bugdayci O, Aribal M E and Akduman I 2025 XGBoost enhances the performance of SAFE: a novel microwave imaging system for early detection of malignant breast cancer Cancers 17 214 [114] Reimer T, Fontaine G and Pistorius S 2025 Interpretable machine learning for tumour detection in microwave breast imaging using image-extracted features 2025 19th European Conf. Antennas and Propagation (EuCAP) (IEEE) pp 1–5 [115] Tariq K, Hassan A, Khan U and Khan W 2025 Breast cancer detection through spectrogram generation from GANs-simulated microwave imaging data 2025 2nd Int. Conf. Microwave, Antennas & Circuits (ICMAC) (IEEE) pp 1–4 [116] Vijayasarveswari V, Mahrom N, Raof R A A, Phak L, Razak M A A, Silan B P, Halim A A, Nasrudin M W, Ramli N and Rahayu Y 2025 Development of statistically modelled feature selection method for microwave breast cancer detection J. Adv. Res. Appl. Sci. Eng. Technol. 50 250–63 [117] Flores-Tapia D and Pistorius S 2011 Real time breast microwave radar image reconstruction using circular holography: a study of experimental feasibility Med. Phys. 38 5420–31 [118] Chew H G, Bogner R E and Lim C C 2001 Dual ν-support vector machine with error rate and training size biasing ICASSP, IEEE Int. Conf. Acoustics, Speech and Signal Processing - Proc. vol 2 (IEEE) pp 1269–72 [119] Zastrow E, Davis S K, Lazebnik M, Kelcz F, Veen B D and Hagness S C 2008 Development of anatomically realistic numerical breast phantoms with accurate dielectric properties for modeling microwave interactions with the human breast IEEE Trans. Biomed. Eng. 55 2792–800 [120] Oliveira B L, O’Halloran M, Conceicao R C, Glavin M and Jones E 2016 Development of clinically informed 3-D tumor models for microwave imaging applications IEEE Antennas Wirel Propag. Lett. 15 520–3 [121] Reimer T, Krenkevich J and Pistorius S 2020 An open-access experimental dataset for breast microwave imaging 2020 14th European Conf. Antennas and Propagation (EuCAP) (IEEE) pp 1–5 [122] Reimer T and Pistorius S 2021 An optimization-based approach to radar image reconstruction in breast microwave sensing Sensors 21 8172 34 Prog. Biomed. Eng. 7(2025) 042008 T M M Silva et al [123] Rocke P, McGinley B, Maher J, Morgan F and Harkin J 2008 Investigating the suitability of FPAAs for evolved hardware spiking neural networks Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) vol 5216 LNCS (Springer) pp 118–29 [124] Salucci M, Vrba J, Merunka I and Massa A 2017 Real-time brain stroke detection through a learning-by-examples technique-an experimental assessment Microw. Opt. Technol. Lett. 59 2796–9 [125] Guo L and Abbosh A 2018 Stroke localization and classification using microwave tomography with k-means clustering and support vector machine Bioelectromagnetics 39 312–24 [126] Saied I M, Arslan T and Chandran S 2022 Classification of Alzheimer’s disease using RF signals and machine learning IEEE J. Electromagn. RF Microw. Med. Biol. 677–85 [127] Hossain A, Islam M T and Almutairi A F 2022 A deep learning model to classify and detect brain abnormalities in portable microwave based imaging system Sci. Rep. 12 6319 [128] Lalitha K and Manjula J 2022 Novel method of characterization of dispersive properties of heterogeneous head tissue using microwave sensing and machine learning algorithms Adv. Electromagn. 11 84–92 [129] Hossain A, Islam M T, Rahim S K A, Rahman M A, Rahman T, Arshad H, Khandakar A, Ayari M A and Chowdhury M E 2023 A lightweight deep learning based microwave brain image network model for brain tumor classification using reconstructed microwave brain (RMB) images Biosensors 13 238 [130] Pokorny T, Vrba J, Fiser O, Vrba D, Drizdal T, Novak M, Tosi L, Polo A and Salucci M 2023 On the role of training data for SVM-based microwave brain stroke detection and classification Sensors 23 2031 [131] Pokorny T, Fiser O, Drizdal T and Vrba J 2023 2D numerical dataset for microwave SVM-based brain stroke classification 2023 Photonics and Electromagnetics Research Symp., PIERS 2023 - Proc. (IEEE) pp 1705–11 [132] Ullah R, Dong Y, Arslan T and Chandran S 2023 A machine learning-based classification method for monitoring Alzheimer’s disease using electromagnetic radar data IEEE Trans. Microw. Theory Tech. 71 4012–26 [133] Hossain A, Islam M T, Rahman T, Chowdhury M E, Tahir A, Kiranyaz S, Mat K, Beng G K and Soliman M S 2023 Brain tumor segmentation and classification from sensor-based portable microwave brain imaging system using lightweight deep learning models Biosensors 13 302 [134] Cardinali L, Spano M, Gugliermino M, Rodriguez-Duarte D O, Ricci M, Vasquez J A T, Palmeri R, Scapaticci R, Crocco L and Vipiana F 2023 A machine learning approach to microwave sensing for non-invasive Alzheimer’s disease early detection 2023 IEEE Int. Conf. Metrology for EXtended Reality, Artificial Intelligence and Neural Engineering (MetroXRAINE) (IEEE) pp 507–12 [135] Taha B, Liza F R, Masud M A, Bepery C, Islam M T and Samsuzzaman M 2023 Brainvisionnet: a deep learning-based approach to evaluate the potential of microwave imaging for classification of brain tumors 2023 Int. Conf. Next-Generation Computing, IoT and Machine Learning (NCIM) (IEEE) pp 1–6 [136] Hasan N, Aktar M, Rana M M and Moni M A 2023 Machine learning-based biomedical antenna for brain tumor detection 2023 Int. Conf. Next-Generation Computing, IoT and Machine Learning (NCIM) (IEEE) pp 1–6 [137] Hasan N, Rana M M, Hasan M M and Moni M A 2023 AI-enhanced biomedical antennas for 2mm brain tumor detection using scattering, admittance and impedance parameters: a comparative analysis 2023 Int. Conf. Information and Communication Technology for Sustainable Development (ICICT4SD) (IEEE) pp 219–23 [138] Hashir M, Khalid N, Mahmood N, Rehman M A, Asad M, Mehmood M Q, Zubair M and Massoud Y 2023 A tinyml based portable, low-cost microwave head imaging system for brain stroke detection 2023 IEEE Int. Symp. on Circuits and Systems (ISCAS) (IEEE) pp 1–4 [139] Cardinali L, Mariano V, Vasquez J T, Crocco L and Vipiana F 2024 Feasibility of Alzheimer’s disease early detection through machine learning applied to microwave sensing data collected from a realistic phantom 2024 IEEE Int. Symp. Antennas and Propagation and INC/USNC-URSI Radio Science Meeting (AP-S/INC-USNC-URSI) (IEEE) pp 241–2 [140] Hossain A, Islam R, Islam M T, Kirawanich P and Soliman M S 2024 FT-FEDTL: a fine-tuned feature-extracted deep transfer learning model for multi-class microwave-based brain tumor classification Comput. Biol. Med. 183 109316 [141] Hasan S, Zamani A, Brankovic A, Bialkowski K S and Abbosh A 2023 Stroke classification with microwave signals using explainable wavelet convolutional neural network IEEE J. Biomed. Health Inf. 28 5667–75 [142] Pokorny T, Vrba D, Fiser O, Salucci M and Vrba J 2024 Systematic optimization of training and setting of SVM-based microwave stroke classification: numerical simulations for 10 port system IEEE J. Electromag. RF Microw. Med. Biol. 8273–81 [143] Lalitha K and Manjula J 2024 Dielectric characterization of dispersive head tissue for detection and classification of tumour using microwave imaging technique and deep learning model Arab. J. Sci. Eng. 49 12305–16 [144] Farhatullah, Chen X, Zeng D, Ullah R, Nawaz R, Xu J and Arslan T 2025 A deep learning approach for non-invasive Alzheimer’s monitoring using microwave radar data Neural Netw. 181 106778 [145] Sudhakaran D U and Bai S T S K 2025 Brain tumor detection using hybrid transfer learning and patch antenna-enhanced microwave imaging Technol. Health Care Official J. Eur. Soc. Eng. Med. 33 9287329251325740 [146] Deebu U S and Sreeja T K 2025 Enhanced brain tumor detection from microwave imaging with hybrid inception-CNN and UWB circular monopole patch antenna Int. J. Intell. Eng. Syst. 18 1216–37 [147] Yuan W, Thammasorn P, Wang L and Mo S 2025 A comprehensive deep learning framework for microwave stroke classification: combining signal analysis, clinical variables and antenna system measurements IEEE Access vol 13, pp 39935–49 [148] Saied I, Arslan T, Chandran S, Smith C, Spires-Jones T and Pal S 2020 Non-invasive RF technique for detecting different stages of Alzheimer’s disease and imaging beta-amyloid plaques and tau tangles in the brain IEEE Trans. Med. Imaging 39 4060–70 [149] Saied I, Bashri M S, Arslan T, Smith C and Chandran S 2019 Dielectric measurements of brain tissues with Alzheimer’s disease pathology in the microwave region Medical Measurements and Applications, MeMeA 2019 - Symp. Proc. (IEEE) pp 1–6 35