SELF-RELIANT RESIDUAL NETWORK BASED DEEP LEARNING FRAMEWORK FOR MELANOMA SKIN DISEASE DETECTION

Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4191 SELF-RELIANT RESIDUAL NETWORK BASED DEEP LEARNING FRAMEWORK FOR MELANOMA SKIN DISEASE DETECTION V. RADHIKA1, A. MUTHUCHUDAR2, M. LINGARAJ3 1Principal, 2 Assistant Professor, Associate Profes sor Department of Computer Science, Sankara College of Science and Commerce, India E-mail: 1 [email protected], [email protected], [email protected] ABSTRACT Melanoma is one of the deadliest types of skin cancer and one of the most aggressive that may be if caught late. While traditional approaches may have their limits, an accurate diagnosis is vital for patient survival. Thanks to its capacity to understand intricate patterns from massive datasets, deep learning has evolved as a potential method for automated melanoma diagnosis. The challenges persist, including overfitting, instability during training, and difficulties in handling nonlinearities, which can hinder accurate predictions. To address these challenges, Self-Reliant ResNet (SR-ResNet) has been proposed. This enhanced version of ResNet integrates Zoutendijk’s Method, a nonlinear optimization technique, to optimize weight updates and improve convergence. SR-ResNet features a series of residual blocks where Zoutendijk’s Method refines the learning process, ensuring stability and efficient training, even in deeper networks. The network’s architecture has been designed to enhance performance and generalization. The proposed SRResNet has been evaluated using a dataset of 10,000 Melanoma Skin Cancer images. The results demonstrate significant improvements in classification accuracy, achieving a high precision rate with reduced overfitting. SR-ResNet outperforms traditional models, establishing itself as a robust tool for melanoma diagnosis. Keywords: Melanoma Skin Cancer, Deep Learning, SR-ResNet, Zoutendijk’s Method, Classification Accuracy 1. INTRODUCTION Melanoma Skin Cancer has emerged as a critical public health concern, recognized for its aggressive nature and high mortality rate if not diagnosed early. This cancer, originating in melanocytes—the cells responsible for producing melanin—demands early detection and precise diagnosis to improve survival rates[1]. Traditional methods, relying on visual examination and biopsy, have exhibited limitations, potentially leading to delays in treatment. The onset of the COVID-19 pandemic has further exacerbated this issue, disrupting routine healthcare services and causing a gradual increase in Melanoma Skin Cancer cases as patients have missed early screenings. This surge has underscored an urgent need for more reliable and advanced diagnostic tools that enhance accuracy and accessibility in the post-pandemic era[2], [3]. Figure 1a shows a benign skin lesion, such as a mole or benign nevus, which is non-cancerous and does not evolve into melanoma. These lesions typically do not invade surrounding tissues or spread to other parts of the body, making them generally harmless.Figure 1b, on the other hand, depicts a malignant skin lesion, specifically melanoma, a dangerous form of skin cancer[4]. Melanoma can grow aggressively, invade nearby tissues, and metastasize to distant organs. Accurate diagnosis and early intervention are crucial for effective treatment of melanoma[4]. Deep learning has gained prominence as a transformative approach in medical imaging, revolutionising how melanoma is detected and classified. Convolutional Neural Networks (CNNs) have emerged as a leading deep learning architecture, showing significant promise in analyzing complex patterns within medical images[5]. Despite these improvements, melanoma detection using deep learning models has been quite difficult. One of the biggest problems is overfitting, which occurs when a model does very well on training data but doesn't adapt well to new, unknown data. A lack of stability in the training process has also been a significant hurdle, often Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4192 resulting in difficulties with convergence in deeper networks. [6], [7]. Figure 1a. Benign Skin Lesion Figure 1b. Malignant Skin Lesion Self-Reliant Residual Network (SRResNet) has been proposed to overcome these challenges, representing an advancement over the traditional ResNet architecture. SR-ResNet integrates Zoutendijk’s Method, a nonlinear optimization technique designed to enhance the training process. While ResNet is known for its powerful architecture, which includes residual connections that mitigate the vanishing gradient problem in deep networks, SR-ResNet builds upon this by optimizing weight updates to improve convergence and stability. IntegratingZoutendijk’s Method ensures that each training step moves the model toward an optimal solution, effectively addressing issues related to overfitting and instability that have previously hindered deeplearning models in melanoma detection. By refining the learning process, SRResNet has established itself as a robust and reliable tool for melanoma detection, offering significant improvements in accuracy and stability. This model’s ability to address the specific challenges posed by melanoma data positions it as a critical advancement in the field, particularly during the COVID-19 pandemic, where the need for early and accurate cancer detection has become increasingly pressing. The healthcare landscape continues to evolve in response to ongoing challenges, with SR-ResNet providing a promising solution for improving patient outcomes in melanoma diagnosis, helping to bridge the gap left by traditional methods and ensuring that more patients receive timely and accurate care. Despite advancements in melanoma detection, existing models suffer from overfitting, unstable convergence, and limited generalization, especially in deeper networks. Ensemble and transfer learning methods offer improvements but lack optimization efficiency and adaptability. Current approaches often overlook nonlinear optimization within residual blocks. This work introduces SR-ResNet, which integrates Zoutendijk’s Method to optimize weight updates, ensuring stable convergence and improved classification accuracy. By addressing a critical gap in convergence-optimized deep learning, SRResNet provides a robust solution tailored for melanoma detection, outperforming traditional models and offering a scalable, precise diagnostic framework suitable for clinical deployment. 2. LITERATURE REVIEW Ensemble Embeddings Approach [8] has explored melanoma detection by integrating an ensemble of machine learning models with deep feature embeddings, demonstrating significant improvements in accuracy. The study highlights the advantages of combining traditional machine learning algorithms with deep learning features, enhancing overall predictive performance. The Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4193 proposed ensemble method mitigates overfitting and increases robustness, making it a promising approach for melanoma detection in clinical settings. MuSClD System [9] has introduced Multisite Cross-organ Calibrated Deep Learning (MuSClD), an automated system for diagnosing non-melanoma skin cancer. The research emphasizes the importance of calibrating deep learning models across multiple sites to improve generalizability. The proposed model leverages cross-organ calibration to enhance the accuracy of diagnosis, addressing the challenges associated with varying data distributions across different clinical settings. Interpretable DL System [10] have developed an interpretable deep learning system for multi-class segmentation and classification of nonmelanoma skin cancer. Their approach combines deep learning with interpretability techniques, allowing clinicians to understand the decisionmaking process of the model. The study demonstrates that interpretable models can achieve high accuracy while providing insights into the model's behavior, which is crucial for clinical adoption. Grasshopper DL Hybrid [11] have proposed a hybrid Grasshopper optimization algorithm combined with deep learning for skin lesion segmentation and melanoma classification. The research integrates evolutionary algorithms with deep learning to optimize the segmentation process, improving classification accuracy. The hybrid approach addresses the limitations of traditional deep learning models in segmenting complex skin lesions, providing a more reliable tool for melanoma diagnosis. Hyperspectral Signature Learning [12]has been explored to classify actinic keratosis and non-melanoma skin cancers using near-infrared hyperspectral signatures. This research shows that hyperspectral imaging in conjunction with machine learning can effectively differentiate between various skin lesions. Their technology provides an alternative to invasive procedures for early identification of skin cancer, which has the potential to greatly improve diagnostic accuracy.Deep Learning Classifier [13]studied the efficacy of several deep learning architectures on datasets consisting of skin lesion information in order to identify and categorize skin cancers. Findings show that deep learning algorithms can detect skin cancer more accurately than conventional approaches, suggesting that they may eventually replace them. It delves into the difficulties encountered by deep learning models, including overfitting and the requirement for extensive datasets. Optimized DL Segmentation [14]highlights the importance of accurate segmentation for better diagnosis accuracy by optimizing deep learning methods for skin lesion segmentation and skin cancer detection. Their research incorporates advanced deep-learning architectures tailored to handle the unique challenges posed by skin lesion images. The study demonstrates that optimized deep-learning models can achieve superior performance in skin cancer detection, particularly in challenging cases. Hybrid DL Framework [15] have introduced a hybrid deep learning framework for predicting skin cancer, combining various deep learning techniques to improve accuracy. The framework integrates convolutional neural networks with other deep learning methods, resulting in a robust model capable of handling diverse skin lesion images. Their approach offers a comprehensive solution for skin cancer prediction, enhancing the model's generalizability across different datasets. Polarimetric ML Classifier [16] has used polarimetric imaging with machine learning to classify non-melanoma skin cancer in mice tissues. The study leverages optical parameters derived from polarimetric imaging to improve classification accuracy. Their findings suggest combining polarimetric imaging with machine learning provides a novel approach to skin cancer diagnosis, potentially offering more accurate results than traditional imaging methods. Transfer Learning Hybrid[17]identify melanoma skin cancer by using transfer learning for segmentation in conjunction with hybrid classification. This study aims to improve melanoma detection systems by utilizing pre-trained models, especially in cases when labeled data is few. The work shows that melanoma diagnosis may be much improved using transfer learning and hybrid classification algorithms, which makes it a viable tool for clinical situations. Ensemble Transfer Learning (ETL) [18]is utilizing deep transfer learning and ensemble machine learning in a combined effort to classify melanoma. In order to improve classification accuracy, the study investigates using deep learning in conjunction with several machine learning models. The ensemble approach addresses the challenges of overfitting and model robustness, providing a more reliable method for melanoma detection. Their findings indicate combining traditional and deep learning methods can lead to Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4194 superior classification performance.DL Metaanalysis (DLMA)[19]have reviewed and analyzed deep learning algorithms for dermoscopy-based melanoma diagnosis in a systematic way. In order to assess deep learning's efficacy in melanoma diagnosis, their study combines results from several investigations. While deep learning models have demonstrated promising results in increasing diagnosis accuracy, the study notes that there is still a need for more research into issues like data quality and model interpretability.Bio-inspired Optimization plays a significant role in different research to achieve the best result [20]-[53]. Even in skin cancer detection also, bio-inspired optimization can be applied. 3. SELF-RELIANT RESIDUAL NETWORK Self-Reliant ResNet (SR-ResNet) represents a substantial advancement over traditional ResNet architectures by integratingZoutendijk's Method, a sophisticated nonlinear optimization technique. By leveraging Zoutendijk's Method, SR-ResNet optimizes the weight update process during training, significantly improving convergence stability and computational efficiency. This integration effectively mitigates issues such as gradient vanishing and overfitting, which are common challenges in deep learning models, particularly in complex tasks like melanoma detection. The enhanced optimization process enables SR-ResNet to achieve more precise and reliable predictions for melanoma. It is a powerful tool in applications that demand high accuracy and robustness in detecting this aggressive form of skin cancer. 3.1. Input Layer The first step in SR-ResNet involves the input layer, where the raw data, typically an image, is fed into the network. The input image can be represented as a tensor of dimensions where represents the height, represents the width and denotes the number of channels, usually corresponding to the color channels in an RGB image . The mathematical formulation of the input tensor can be described as: for (1) In the initial convolutional layer, the input tensor undergoes a convolution operation with a set of filters of kernels, denoted as , where is a four-dimensional tensor of dimensions Here, and denote the height and width of the filter, respectively, and represents the number of filters applied. The convolution operation between and can be mathematically represented as: (2) where is the output feature map resulting from the convolution operation at position . The above equation computes the weighted sum of the input pixel values and the corresponding filter weights, producing an output feature map of reduced dimensions, which captures the essential features from the input image. A nonlinearity is introduced into the model after the convolution process by applying an activation function element-wise to the output feature map . The mathematical expression of the most used activation function, the Rectified Linear Unit (ReLU): (3) where is the activated output at position . This operation retains positive values and sets all negative values to zero, which helps to prevent the vanishing gradient problem. In specific architectures, a max-pooling operation follows the activation function, reducing the spatial dimensions of the feature maps by selecting the maximum value from non-overlapping subregions within the feature map. This operation can be defined as: for (4) where is the pooled output, represents the size of the pooling window, and is the activated feature map. This pooling process reduces the computational complexity and provides spatial invariance, thereby setting up the initial feature map for further processing in the subsequent layers of SR-ResNet. 3.2. Initial Convolution and Pooling The initial feature map undergoes further processing through convolutional layers within the residual block. The initial feature map obtained after the first convolutional and pooling operations, is passed through a series of convolutional layers designed to extract deeper features. Let have dimensions. The first convolutional Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4195 layer within this block applies a set of filters. with dimensions where represents the number of filters. The operation is defined as: (5 ) where represents the output feature map after the first convolutional operation in layer . A batch normalization step is applied to the output feature map. , which normalizes the output to have a mean of zero and a variance of one. Let and represent the mean and variance, respectively: (6) where is a small constant added for numerical stability. Following batch normalization, the activation function , typically ReLU, is applied: (7 ) This process is repeated in subsequent convolutional layers within the same residual block. Consider the next convolutional layer with filters. of dimensions , where is the number of filters in the second convolutional layer. The output of this layer is: (8 ) Batch normalization is again applied to (9) Following normalization, the ReLU activation function is applied: (10 ) In a standard residual block, these operations are followed by adding the original input to the final output of the convolutional layers. In SR-ResNet, this step incorporates Zoutendijk's Method for optimization.The residual function is denoted as: (11) The input is added to the residual function. yielding the output of the residual block: (12) Zoutendijk's Method is then applied to optimize , where the feasible direction and step size are calculated. The update step for the parameters involves moving in the direction by a step size (13) The optimized output becomes the input to the next block in the SR-ResNet, ensuring efficient training and enhanced convergence throughout the network. 3.3. Residual Block with Zoutendijk's Optimization The process continues with stacking multiple residual blocks, each incorporating the enhanced optimization technique inspired by Zoutendijk's Method. The input to each subsequent residual block is the output from the previous block, optimized to improve training stability and convergence. Let represent the output of the residual block, which serves as the input to the next block.The input is first passed through a convolutional layer with filters defined by the dimensions where is the number of channels in and is the number of filters. The operation is expressed as: (14 ) Following this convolution, the batch normalization step is applied to the output feature map. with the mean and variance calculated as: (15) The output is then passed through the activation function, typically ReLU: (16 ) Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4196 The process continues with another convolutional layer, this time using filters. with dimensions where represents the number of filters in the next layer. The convolution operation is given by: (1 7) Batch normalization is again applied: (18) The activated output is then: (19 ) This output serves as the residual function for the next layer. To maintain continuity, the original input is added to this residual function, forming the input to the next block: (20) At this stage, Zoutendijk's Method is applied again. The objective function representing the loss, is optimized by determining a feasible direction and an optimal step size (21) The optimized set of parameters updates the weights within the block, ensuring that the network continues to learn efficiently. The output becomes the input for the subsequent residual block, preserving the gradient flow and enhancing the overall network performance.By repeating this process, SR-ResNet constructs a deep network where each residual block is optimized for improved convergence, leading to better model accuracy and efficiency in learning complex patterns. 3.4. Stacking SR-ResNet Blocks The output from the series of residual blocks is processed through global average pooling, fully connected layers, and the final optimization steps to produce the network's output. The input to this step is the output from the last residual block, denoted as where is the index of the final residual block in the network.The global average pooling operation is applied to which has dimensions The feature maps are aggregated into a single value using the global average pooling method, which averages all the spatial components.This operation is mathematically expressed as: (22) where is the pooled output corresponding to the feature map, and The result is a vector of length. This pooled vector is then passed to a fully connected layer with a weight matrix and bias vector The fully connected layer computes the output as: (23) where is the output of the neuron in the fully connected layer, and with being the number of neurons in the fully connected layer. After the fully connected layer, the outputs are transformed into probabilities using a softmax activation function. The softmax function for the th class is defined as: (24) where represents the predicted probability of the class. The loss function is determined during training by using cross-entropy loss, which quantifies the discrepancy between the actual labels and the anticipated probabilities .The crossentropy loss is given by: (25) where is the ground truth label for the class. To optimize the network, Zoutendijk's Method is applied, which involves determining the feasible direction and the optimal step size for minimizing the loss function. The weight updates for the fully connected layer are performed as: (26) Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4197 where represents the weights at iteration and is the feasible direction determined by Zoutendijk's Method. The back propagation process updates the parameters throughout the network, including the weights in the convolutional layers and fully connected layers, ensuring the optimization of the overall objective function The final output of SR-ResNet is the class with the highest probability from the softmax layer, representing the network's prediction. 3.5. Bottleneck Blocks The primary focus is optimizing the entire network using Zoutendijk's Method, which is integrated with the backpropagation algorithm. The optimization process begins with calculating gradients for all network parameters, followed by determining feasible directions and optimal step sizes for weight updates.To initiate the optimization, the loss function where represents the entire set of network parameters, is computed. Each parameter is used to represent the gradient of the loss function. (27) The chain rule is used to compute the gradient. for every layer in the network. To minimize the loss, the parameters might be modified in the direction indicated by this gradient.The gradient at layer for weight is given by: (28) The output of the layer before adding an activation function is represented by .Once the gradients are calculated, Zoutendijk's Method is applied to determine a feasible direction. for each parameter update. This involves solving the optimization problem: (29) (30) where represents any constraints on the parameters and is the step size. The direction is selected to ensure that the loss decreases while maintaining feasibility concerning the constraints. The optimal step size is determined by minimizing the loss along the direction : (31) The parameters for layer, are then updated using the feasible direction and the calculated step size as follows: (32) This update rule is applied to all layers, ensuring that each set of parameters is optimized according to Zoutendijk's Method. The process is iterative, with gradients recalculated after each update until convergence is achieved, i.e., when the loss function stabilizes at a minimum or nearminimum value.The back propagation process, in conjunction with Zoutendijk's Method, ensures that the network's parameters are updated in the direction that reduces the loss and optimally within the feasible region defined by any constraints. This results in a robust and efficient learning process for the entire SR-ResNet architecture. 3.6. Downsampling The focus is on the iterative process of refining the model through repeated application of Zoutendijk's Method across multiple training epochs. The process ensures that the model progressively moves closer to the global or local minimum of the loss function while respecting any constraints imposed on the parameters. Given the parameter set at epoch the objective is to further minimize the loss function by updating the parameters through a series of iterations. To get the loss function's gradient concerning the parameters, one uses the following formula: (33) Zoutendijk's Method is then applied to determine the feasible direction. that minimizes the objective function while ensuring the parameters remain within the feasible region. This optimization problem can be expressed as: (34) (35) Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4198 where represents any constraints, such as nonnegativity or boundedness, that must be satisfied by the parameters . To refine the parameters, the optimal step size is determined by minimizing the loss along the direction : (36) The parameters are then updated for the next iteration using the calculated direction and step size: (37) As training progresses through multiple epochs, the gradient is recalculated after each parameter update, ensuring that the parameters continually move in the direction most effectively reduces the loss function.The iteration process continues with updated gradients. and directions for each subsequent epoch: (38) (39) The refinement process also involves ensuring that the constraints remain satisfied. This requires recalculating the feasible region at each iteration and adjusting the step size. accordingly to maintain feasibility: (40) This iterative process continues across multiple epochs until the loss function reaches a stable minimum, indicating that the model parameters have converged to optimal values. The repeated application of Zoutendijk's Method throughout this process ensures that the parameter updates are effective in reducing the loss and robust against potential constraint violations, leading to a well-optimized model in the SR-ResNet framework. 3.7. Global Average Pooling The focus is on evaluating and adjusting the model during the training process to ensure the convergence and stability of the optimization. This step involves monitoring the loss function, changing the learning rate, and applying regularization techniques to enhance the model's generalization capabilities.The primary objective in this step is to minimize the loss function by refining the parameters while ensuring that the model does not overfit the training data. The process begins by evaluating the loss function after each training epoch. The loss function at epoch is denoted as: (41 ) where represents the number of training samples, is the true label, and is predicted probability for the sample. The loss function's gradient concerning the parameters may be determined using the following formula: (42 ) Applying regularization methods, such as L2 regularization, helps avoid overfitting and enhances generalization.The L2 regularization term is added to the loss function: (43) where controls the penalty's intensity as a regularization parameter. The total loss with regularization is minimized by updating the parameters using the gradient descent method with a learning rate. that is potentially adjusted during training: (44) To ensure stable convergence, the learning rate may be adjusted based on the evaluation of the loss function over successive epochs. If the loss does not decrease as expected, the learning rate is reduced: with (45) This step also involves monitoring the model's performance on a validation set. The validation loss is calculated similarly to the training loss. If the validation loss starts increasing while the training loss decreases, early stopping criteria may be applied to prevent overfitting: Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4199 If , then stop training (46) Continuous evaluation of the loss function, adjustment of the learning rate, and application of regularization ensure that SR-ResNet converges to a well-generalized solution that effectively performs on training and unseen data. 3.8. Fully Connected Layer In Fully Connected Layer, the process involves final adjustments and the evaluation of the trained model's performance on the test data. This step ensures the model generalizes well to new, unseen data and evaluates its robustness across different scenarios.The trained model parameters obtained after the training process, are used to make predictions on the test dataset. For each test sample the model computes the predicted probability for the target class using the softmax function: (47) where denotes the number of classes and represents the output of the final layer for the test sample. The predicted class label for each test sample is determined by selecting the class with the highest predicted probability: (48) Next, measures like F1-score, recall, accuracy, and precision assess the model's overall performance. Using the number of test samples divided by the number of properly predicted labels, we can determine the accuracy: (49) The indicator function is defined as 1 when the predicted label is identical to the real label and 0 otherwise, with being the total number of test samples. Precision, recall, and F1-score are calculated to provide a more detailed analysis of the model's performance, particularly in imbalanced datasets. Precision for a class is defined as: (50) where and denote the true positives and false positives for class , respectively. Recall for class is defined as: (51) where denotes the false negatives for class . The F1-score for class is the harmonic mean of precision and recall: (52) To assess the overall performance across all classes, the macro-averaged F1-score is calculated: (53) After evaluating the performance on the test set, if the model's performance metrics meet the desired criteria, the model is considered ready for deployment. However, if the performance is suboptimal, the training process may be revisited, and adjustments to the learning rate, regularization, or architecture may be made. 3.9. Softmax Activation SR-ResNet undergoes refinement through fine-tuning, which aims to optimize the network's performance on a specific task or dataset. Finetuning involves retraining some or all network layers using a lower learning rate while potentially adjusting other hyper parameters to enhance performance. The fine-tuning process begins by reevaluating the loss function on the target dataset, where represents the current set of network parameters. A new calculation is made for the loss function's gradient concerning these parameters, and it is: (54) In fine-tuning, the learning rate is typically reduced compared to the initial training phase to allow for more precise weight adjustments. The learning rate at iteration during fine-tuning is expressed as: (55) where is the initial learning rate used during finetuning, and is a decay factor such that Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4206 AI, and optimizing deployment for clinical integration. REFERENCES [1] K. Sethanan et al., “Double AMIS-ensemble deep learning for skin cancer classification,” Expert Syst. Appl., vol. 234, p. 121047, 2023, doi: https://doi.org/10.1016/j.eswa.2023.121047. [2] W. Abbes and D. Sellami, “Deep Neural Networks for Melanoma Detection from Optical Standard Images using Transfer Learning,” Procedia Comput. Sci., vol. 192, pp. 1304–1312, 2021, doi: https://doi.org/10.1016/j.procs.2021.08.134. [3] F. Alenezi, A. Armghan, and K. Polat, “A multi-stage melanoma recognition framework with deep residual neural network and hyperparameter optimization-based decision support in dermoscopy images,” Expert Syst. Appl., vol. 215, p. 119352, 2023, doi: https://doi.org/10.1016/j.eswa.2022.119352. [4] S. Kadry, E. Verdú, R. Damasevicius, L. Abualigah, V. Singh, and V. Rajinikanth, “CNN segmentation of skin melanoma in preprocessed dermoscopy images,” Procedia Comput. Sci., vol. 235, pp. 2775–2782, 2024, doi: https://doi.org/10.1016/j.procs.2024.04.262. [5] Q. Liu, H. Kawashima, and A. Rezaei sofla, “An optimal method for melanoma detection from dermoscopy images using reinforcement learning and support vector machine optimized by enhanced fish migration optimization algorithm,” Heliyon, vol. 9, no. 10, p. e21118, 2023, doi: https://doi.org/10.1016/j.heliyon.2023.e21118. [6] L. Andersson et al., “Poly(ethylene glycol)- poly(ester-carbonate) block copolymers carrying PEG-peptidyl-doxorubicin pendant side chains: Synthesis and evaluation as anticancer conjugates,” Biomacromolecules, vol. 6, no. 2, pp. 914–926, 2005, doi: 10.1021/bm049381p. [7] Ç. Suiçmez, H. Tolga Kahraman, A. Suiçmez, C. Yılmaz, and F. Balcı, “Detection of melanoma with hybrid learning method by removing hair from dermoscopic images using image processing techniques and wavelet transform,” Biomed. Signal Process. Control, vol. 84, p. 104729, 2023, doi: 10.1016/j.bspc.2023.104729. [8] S. Ghosh, S. Dhar, R. Yoddha, S. Kumar, A. K. Thakur, and N. D. Jana, “Melanoma Skin Cancer Detection Using Ensemble of Machine Learning Models Considering Deep Feature Embeddings,” Procedia Comput. Sci., vol. 235, pp. 3007–3015, 2024, doi: https://doi.org/10.1016/j.procs.2024.04.284. [9] Y. Zhou et al., “Multi-site cross-organ calibrated deep learning (MuSClD): Automated diagnosis of non-melanoma skin cancer,” Med. Image Anal., vol. 84, p. 102702, 2023, doi: https://doi.org/10.1016/j.media.2022.102702. [10] S. M. Thomas, J. G. Lefevre, G. Baxter, and N. A. Hamilton, “Interpretable deep learning systems for multi-class segmentation and classification of non-melanoma skin cancer,” Med. Image Anal., vol. 68, p. 101915, 2021, doi: https://doi.org/10.1016/j.media.2020.101915. [11] P. Thapar, M. Rakhra, M. Alsaadi, A. Quraishi, A. Deka, and J. V. Naga Ramesh, “A hybrid Grasshopper optimization algorithm for skin lesion segmentation and melanoma classification using deep learning,” Healthc. Anal., vol. 5, p. 100326, 2024, doi: https://doi.org/10.1016/j.health.2024.100326. [12] L. A. Courtenay, I. Barbero-García, S. Martínez-Lastras, S. Del Pozo, M. Corral, and D. González-Aguilera, “Using computational learning for non-melanoma skin cancer and actinic keratosis near-infrared hyperspectral signature classification,” Photodiagnosis Photodyn. Ther., vol. 49, p. 104269, 2024, doi: https://doi.org/10.1016/j.pdpdt.2024.104269. [13] C. Kavitha, S. Priyanka, M. P. Kumar, and V. Kusuma, “Skin Cancer Detection and Classification using Deep Learning Techniques,” Procedia Comput. Sci., vol. 235, pp. 2793–2802, 2024, doi: https://doi.org/10.1016/j.procs.2024.04.264. [14] R. R. Babu and F. M. Philip, “Optimized deep learning for skin lesion segmentation and skin cancer detection,” Biomed. Signal Process. Control, vol. 95, p. 106292, 2024, doi: https://doi.org/10.1016/j.bspc.2024.106292. [15] E. Farea, R. A. A. Saleh, H. AbuAlkebash, A. A. R. Farea, and M. A. Al-antari, “A hybrid deep learning skin cancer prediction framework,” Eng. Sci. Technol. an Int. J., vol. 57, p. 101818, 2024, doi: https://doi.org/10.1016/j.jestch.2024.101818. [16] T.-T.-H. Pham, T.-N. Luu, T.-V. Nguyen, N.- T. Huynh, Q.-H. Phan, and T.-H. Le, “Polarimetric imaging combining optical parameters for classification of mice nonmelanoma skin cancer tissue using machine Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4207 learning,” Heliyon, vol. 9, no. 11, p. e22081, 2023, doi: https://doi.org/10.1016/j.heliyon.2023.e22081. [17] R. Dandu, M. Vinayaka Murthy, and Y. B. Ravi Kumar, “Transfer learning for segmentation with hybrid classification to Detect Melanoma Skin Cancer,” Heliyon, vol. 9, no. 4, p. e15416, 2023, doi: https://doi.org/10.1016/j.heliyon.2023.e15416. [18] M. Roshni Thanka et al., “A hybrid approach for melanoma classification using ensemble machine learning techniques with deep transfer learning,” Comput. Methods Programs Biomed. Updat., vol. 3, p. 100103, 2023, doi: https://doi.org/10.1016/j.cmpbup.2023.100103. [19] Z. Ye et al., “Deep learning algorithms for melanoma detection using dermoscopic images: A systematic review and metaanalysis,” Artif. Intell. Med., vol. 155, p. 102934, 2024, doi: https://doi.org/10.1016/j.artmed.2024.102934. [20] B. Suchitra, J. Ramkumar, and R. Karthikeyan, “Frog Leap Inspired Optimization-Based Extreme Learning Machine For Accurate Classification Of Latent Autoimmune Diabetes In Adults (LADA),” J. Theor. Appl. Inf. Technol., vol. 103, no. 2, pp. 472–494, 2025, [Online]. Available: https://www.scopus.com/inward/record.uri?eid =2-s2.085217140979&partnerID=40&md5=9540433c 16d5ff0f6c2de4b8c43a4812 [21] J. Ramkumar, V. Valarmathi, and R. Karthikeyan, “Optimizing Quality of Service and Energy Efficiency in Hazardous Drone Ad-Hoc Networks (DANET) Using Kingfisher Routing Protocol (KRP),” Int. J. Eng. Trends Technol., vol. 73, no. 1, pp. 410–430, 2025, doi: 10.14445/22315381/IJETT-V73I1P135. [22] R. Jaganathan, S. Mehta, and R. Krishan, “Preface,” Bio-Inspired Intell. Smart Decis., pp. xix–xx, 2024, [Online]. Available: https://www.scopus.com/inward/record.uri?eid =2-s2.085195725049&partnerID=40&md5=7a2aa7adc 005662eebc12ef82e3bd19f [23] J. Ramkumar, B. Varun, V. Valarmathi, D. R. Medhunhashini, and R. Karthikeyan, “JaguarBased Routing Protocol (Jrp) For Improved Reliability And Reduced Packet Loss In Drone Ad-Hoc Networks (DANET),” J. Theor. Appl. Inf. Technol., vol. 103, no. 2, pp. 696–713, 2025, [Online]. Available: https://www.scopus.com/inward/record.uri?eid =2-s2.085217213044&partnerID=40&md5=e38a375e 46cf43c95d6702a3585a7073 [24] J. Ramkumar and R. Vadivel, “Multi-Adaptive Routing Protocol for Internet of Things based Ad-hoc Networks,” Wirel. Pers. Commun., vol. 120, no. 2, pp. 887–909, Apr. 2021, doi: 10.1007/s11277-021-08495-z. [25] A. Senthilkumar, J. Ramkumar, M. Lingaraj, D. Jayaraj, and B. Sureshkumar, “Minimizing Energy Consumption in Vehicular Sensor Networks Using Relentless Particle Swarm Optimization Routing,” Int. J. Comput. Networks Appl., vol. 10, no. 2, pp. 217–230, 2023, doi: 10.22247/ijcna/2023/220737. [26] S. P. Priyadharshini and J. Ramkumar, “Mappings Of Plithogenic Cubic Sets,” Neutrosophic Sets Syst., vol. 79, pp. 669–685, 2025, doi: 10.5281/zenodo.14607210. [27] J. Ramkumar and R. Vadivel, “Improved Wolf prey inspired protocol for routing in cognitive radio Ad Hoc networks,” Int. J. Comput. Networks Appl., vol. 7, no. 5, pp. 126–136, 2020, doi: 10.22247/ijcna/2020/202977. [28] R. Jaganathan and R. Vadivel, “Intelligent Fish Swarm Inspired Protocol (IFSIP) for Dynamic Ideal Routing in Cognitive Radio Ad-Hoc Networks,” Int. J. Comput. Digit. Syst., vol. 10, no. 1, pp. 1063–1074, 2021, doi: 10.12785/ijcds/100196. [29] J. Ramkumar and R. Vadivel, “CSIP—cuckoo search inspired protocol for routing in cognitive radio ad hoc networks,” in Advances in Intelligent Systems and Computing, Springer Verlag, 2017, pp. 145–153. doi: 10.1007/978981-10-3874-7_14. [30] S. P. Geetha, N. M. S. Sundari, J. Ramkumar, and R. Karthikeyan, “Energy Efficient Routing In Quantum Flying Ad Hoc Network (Q-Fanet) Using Mamdani Fuzzy Inference Enhanced Dijkstra’s Algorithm (MFI-EDA),” J. Theor. Appl. Inf. Technol., vol. 102, no. 9, pp. 3708– 3724, 2024, [Online]. Available: https://www.scopus.com/inward/record.uri?eid =2-s2.085197297302&partnerID=40&md5=72d51668 bee6239f09a59d2694df67d6 [31] M. P. Swapna, J. Ramkumar, and R. Karthikeyan, “Energy-Aware Reliable Routing with Blockchain Security for Heterogeneous Wireless Sensor Networks,” in Lecture Notes in Networks and Systems, V. Goar, M. Kuri, R. Kumar, and T. Senjyu, Eds., Springer Science and Business Media Deutschland GmbH, 2025, pp. 713–723. doi: 10.1007/978-981-97-61067_43. Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4208 [32] J. Ramkumar, K. S. Jeen Marseline, and D. R. Medhunhashini, “Relentless Firefly Optimization-Based Routing Protocol (RFORP) for Securing Fintech Data in IoTBased Ad-Hoc Networks,” Int. J. Comput. Networks Appl., vol. 10, no. 4, pp. 668–687, 2023, doi: 10.22247/ijcna/2023/223319. [33] K. S. J. Marseline, J. Ramkumar, and D. R. Medhunhashini, “Sophisticated Kalman Filtering-Based Neural Network for Analyzing Sentiments in Online Courses,” in Smart Innovation, Systems and Technologies, A. K. Somani, A. Mundra, R. K. Gupta, S. Bhattacharya, and A. P. Mazumdar, Eds., Springer Science and Business Media Deutschland GmbH, 2024, pp. 345–358. doi: 10.1007/978-981-97-3690-4_26. [34] J. Ramkumar and R. Vadivel, “Whale optimization routing protocol for minimizing energy consumption in cognitive radio wireless sensor network,” Int. J. Comput. Networks Appl., vol. 8, no. 4, pp. 455–464, 2021, doi: 10.22247/ijcna/2021/209711. [35] M. P. Swapna and J. Ramkumar, “Multiple Memory Image Instances Stratagem to Detect Fileless Malware,” in Communications in Computer and Information Science, S. Rajagopal, K. Popat, D. Meva, and S. Bajeja, Eds., Springer Science and Business Media Deutschland GmbH, 2024, pp. 131–140. doi: 10.1007/978-3-031-59100-6_11. [36] D. Jayaraj, J. Ramkumar, M. Lingaraj, and B. Sureshkumar, “AFSORP: Adaptive Fish Swarm Optimization-Based Routing Protocol for Mobility Enabled Wireless Sensor Network,” Int. J. Comput. Networks Appl., vol. 10, no. 1, pp. 119–129, 2023, doi: 10.22247/ijcna/2023/218516. [37] M. Lingaraj, T. N. Sugumar, C. S. Felix, and J. Ramkumar, “Query aware routing protocol for mobility enabled wireless sensor network,” Int. J. Comput. Networks Appl., vol. 8, no. 3, pp. 258–267, 2021, doi: 10.22247/ijcna/2021/209192. [38] R. Jaganathan, S. Mehta, and R. Krishan, BioInspired intelligence for smart decisionmaking. IGI Global, 2024. doi: 10.4018/9798369352762. [39] R. Vadivel and J. Ramkumar, “QoS-enabled improved cuckoo search-inspired protocol (ICSIP) for IoT-based healthcare applications,” in Incorporating the Internet of Things in Healthcare Applications and Wearable Devices, IGI Global, 2019, pp. 109–121. doi: 10.4018/978-1-7998-1090-2.ch006. [40] J. Ramkumar, R. Karthikeyan, and M. Lingaraj, “Optimizing IoT-Based Quantum Wireless Sensor Networks Using NM-TEEN Fusion of Energy Efficiency and Systematic Governance,” in Lecture Notes in Electrical Engineering, V. Shrivastava, J. C. Bansal, and B. K. Panigrahi, Eds., Springer Science and Business Media Deutschland GmbH, 2025, pp. 141–153. doi: 10.1007/978-981-97-6710-6_12. [41] J. Ramkumar, R. Karthikeyan, and V. Valarmathi, “Alpine Swift Routing Protocol (ASRP) for Strategic Adaptive Connectivity Enhancement and Boosted Quality of Service in Drone Ad Hoc Network (DANET),” Int. J. Comput. Networks Appl., vol. 11, no. 5, pp. 726–748, 2024, doi: 10.22247/ijcna/2024/45. [42] P. Menakadevi and J. Ramkumar, “Robust Optimization Based Extreme Learning Machine for Sentiment Analysis in Big Data,” in 2022 International Conference on Advanced Computing Technologies and Applications, ICACTA 2022, Institute of Electrical and Electronics Engineers Inc., 2022. doi: 10.1109/ICACTA54488.2022.9753203. [43] J. Ramkumar, A. Senthilkumar, M. Lingaraj, R. Karthikeyan, and L. Santhi, “OPTIMAL APPROACH FOR MINIMIZING DELAYS IN IOT-BASED QUANTUM WIRELESS SENSOR NETWORKS USING NM-LEACH ROUTING PROTOCOL,” J. Theor. Appl. Inf. Technol., vol. 102, no. 3, pp. 1099–1111, 2024, [Online]. Available: https://www.scopus.com/inward/record.uri?eid =2-s2.085185481011&partnerID=40&md5=bf0ff974c eabc0ad58e589b28797c684 [44] J. Ramkumar and R. Vadivel, “Improved frog leap inspired protocol (IFLIP) – for routing in cognitive radio ad hoc networks (CRAHN),” World J. Eng., vol. 15, no. 2, pp. 306–311, 2018, doi: 10.1108/WJE-08-2017-0260. [45] R. Jaganathan, S. Mehta, and R. Krishan, Intelligent Decision Making Through BioInspired Optimization. IGI Global, 2024. doi: 10.4018/979-8-3693-2073-0. [46] R. Jaganathan and V. Ramasamy, “Performance modeling of bio-inspired routing protocols in Cognitive Radio Ad Hoc Network to reduce end-to-end delay,” Int. J. Intell. Eng. Syst., vol. 12, no. 1, pp. 221–231, 2019, doi: 10.22266/IJIES2019.0228.22. [47] R. Karthikeyan and R. Vadivel, “Boosted Mutated Corona Virus Optimization Routing Protocol (BMCVORP) for Reliable Data Transmission with Efficient Energy Journal of Theoretical and Applied Information Technology 31st May 2025. Vol.103. No.10 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4209 Utilization,” Wirel. Pers. Commun., vol. 135, no. 4, pp. 2281–2301, 2024, doi: 10.1007/s11277-024-11155-7. [48] R. Karthikeyan and R. Vadivel, “Proficient Dazzling Crow Optimization Routing Protocol (PDCORP) for Effective Energy Administration in Wireless Sensor Networks,” in IEEE International Conference on Electrical, Electronics, Communication and Computers, ELEXCOM 2023, 2023, pp. 1–6. doi: 10.1109/ELEXCOM58812.2023.10370559. [49] N. K. Ojha, A. Pandita, and J. Ramkumar, “Cyber security challenges and dark side of AI: Review and current status,” in Demystifying the Dark Side of AI in Business, 2024, pp. 117– 137. doi: 10.4018/979-8-3693-0724-3.ch007. [50] J. Ramkumar, S. S. Dinakaran, M. Lingaraj, S. Boopalan, and B. Narasimhan, “IoT-Based Kalman Filtering and Particle Swarm Optimization for Detecting Skin Lesion,” in Lecture Notes in Electrical Engineering, K. Murari, N. Prasad Padhy, and S. Kamalasadan, Eds., Singapore: Springer Nature Singapore, 2023, pp. 17–27. doi: 10.1007/978-981-198353-5_2. [51] J. Ramkumar, C. Kumuthini, B. Narasimhan, and S. Boopalan, “Energy Consumption Minimization in Cognitive Radio Mobile AdHoc Networks using Enriched Ad-hoc Ondemand Distance Vector Protocol,” in 2022 International Conference on Advanced Computing Technologies and Applications, ICACTA 2022, 2022. doi: 10.1109/ICACTA54488.2022.9752899. [52] J. Ramkumar, R. Vadivel, and B. Narasimhan, “Constrained Cuckoo Search Optimization Based Protocol for Routing in Cloud Network,” Int. J. Comput. Networks Appl., vol. 8, no. 6, pp. 795–803, 2021, doi: 10.22247/ijcna/2021/210727. [53] L. Mani, S. Arumugam, and R. Jaganathan, “Performance Enhancement of Wireless Sensor Network Using Feisty Particle Swarm Optimization Protocol,” ACM Int. Conf. Proceeding Ser., pp. 1–5, Dec. 2022, doi: 10.1145/3590837.3590907. [54] M. H. Javid, “Melanoma Skin Cancer Dataset of 10000 Images.” Kaggle, 2022. doi: 10.34740/KAGGLE/DSV/3376422.