Alzheimer Classifications Combining Machine Learning and Signal Processing

International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249-8958 (Online), Volume-15 Issue-1, October 2025 18 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.E467314050625 DOI: 10.35940/ijeat.E4673.15011025 Journal Website: www.ijeat.org Abstract: Manual detection of presence of Alzheimer even determination of its intensity from Magnetic Resonance Imaging (MRI) image is an easy task when a patient is heavily affected by the disease. The situation becomes cumbersome for a physician when a very mild affected image comes under consideration. For cases involving subtle differences in images, Machine Learning (ML) and Deep Learning (DL) classification are considered the best solutions. In this paper, Alzheimer-affected images of four categories —mild impairment, moderate impairment, very mild impairment, and no impairment — are taken from a benchmark open database. The image set is converted to a numerical feature vector using Speeded-Up Robust Features (SURF), Features from Accelerated Segment Test (FAST), Harris–Stephen’s, and minimum eigenvalue. The feature vector is then applied to several ML algorithms: Fuzzy Inference System (FIS), Multiple Linear Regression (MLR), Fuzzy C-Means (FCM), Naïve Bayes and linear Support Vector Machine (SVM) to check their eligibility in classifying these four types of medical images. The accuracy of detection from any individual method ranges from 56% to 68%. However, applying the maximum voting scheme to all the MLs yields an accuracy of 75%. The long feature vector is also extracted from the discrete wavelet transform (DWT), and classification is done from its coherence, but the accuracy is found to be very poor. Next, the image dataset is applied to the Bag of Features (BoF) algorithm using 500 visual words, yielding a moderate result with an accuracy of 84.28%. Finally, two deep learning models are applied: PyTorch Convolutional Neural Network (CNN) and Keras Visual Geometry Group (VGG) 19. These models are tested on 20% of the trained images, achieving accuracies of 95.7% and 97.1%, respectively. Keywords: BoF, Accuracy, Coherence of DWT, FIS, and Parallel Plot. Abbreviations: AD: Alzheimer's disease CNNs: Convolutional Neural Networks MCI: Mild Cognitive Impairment Manuscript received on 23 May 2025 | First Revised Manuscript received on 30 May 2025 | Second Revised Manuscript received on 17 September 2025 | Manuscript Accepted on 15 October 2025 | Manuscript published on 30 October 2025. *Correspondence Author(s) Samsi Ara*, Department of Computer Science & Engineering, Jahangirnagar University, Savar (Dhaka), Bangladesh. Email ID: nahar495[email protected]om, ORCID ID: 0009-0002-5306-4940 Md. Imdadul Islam, Department of Computer Science & Engineering, Jahangirnagar University, Savar (Dhaka), Bangladesh. Email ID: [email protected], ORCID ID: 0000-0003-2045-6382 Jugal Krishna Das, Department of Computer Science & Engineering, Jahangirnagar University, Savar (Dhaka), Bangladesh. Email ID: [email protected], ORCID ID: 0009-0000-8572-9375 Md. Golam Saklayen, Department of Applied Physics and Electronic Engineering, Rajshahi University, Rajshah (Dhaka), Bangladesh. Email ID: [email protected], ORCID ID: 0000-0001-5230-0089 Md. Mizanur Rahman, Department of Applied Physics and Electronic Engineering, Rajshahi University, Rajshahi (Dhaka), Bangladesh. Email ID: mizan.apee[email protected], ORCID ID: 0009-0009-4639-247X © The Authors. Published by Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open-access article under the CC-BY-NC-ND license http://creativecommons.org/licenses/by-nc-nd/4.0/ MLR: Multiple Linear Regression CN: Cognitively Normal MRI: Magnetic Resonance Imaging VMD: Very Mild Dementia MoD: Moderate Dementia LSTM: Long Short-Term Memory DL: Deep Learning ML: Machine Learning SURF: Speeded-Up Robust Features FIS: Fuzzy Inference System BoF: Bag of Features I. INTRODUCTION The brain is considered one of the most pivotal organs in the human body. All the activities and responses that allow us to think and believe are administered and facilitated by the brain. It also empowers our attitude and recollections. Alzheimer's Disease (AD) is an incurable, progressive neurological brain disorder (eg, memory loss). It happens when plaques containing beta amyloid protein form in the brain. Every three seconds, someone in the world is diagnosed with Alzheimer's disease. According to the World Alzheimer’s report statistics, People over the age of 65 are most likely to get this condition. More than 55 million people have been diagnosed with Alzheimer’s disease, and this number is increasing day by day, with 78 million expected by 2030, as shown in [1]. The patient will forget current events in the early stages. If the illness progresses, they will moderately forget whole events. It is indispensable to diagnose the disease as soon as possible. Initial detection of Alzheimer's disease can help with proper treatment and prevent brain tissue damage. There is no cure for this condition, but modern medications and treatments are available to slow down the disease progression temporarily. AD can be more precisely categorised using segmented MRI scans due to in-depth analyses of tissue architecture. Several complex segmentation approaches have been proposed for Alzheimer's disease (AD) diagnosis. Since machine learning algorithms can yield effective results over an extensive data collection, they have received interest for use in segmenting the brain’s structure and classifying AD. The rest of the paper is organized as: section II discusses about results of some recent works relevant to detection of levels of Alzheimer’s affect, section III deals with methodology of this research work contains image processing, ML and DL to classify disease affected images, section IV provides the results using the method of the paper and section V concludes the entire paper with some future recommendations. Alzheimer Classifications Combining Machine Learning and Signal Processing Samsi Ara, Md. Imdadul Islam, Jugal Krishna Das, Md. Golam Saklayen, Md. Mizanur Rahman Alzheimer Classifications Combining Machine Learning and Signal Processing 19 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.E467314050625 DOI: 10.35940/ijeat.E4673.15011025 Journal Website: www.ijeat.org II. LITERATURE REVIEW This section discusses some state-of-the-art works pertinent to image classification, specifically the identification of diseases from medical images. AD is the most common form of dementia. It typically manifests as a progressive decline in cognitive function and memory, subsequently impairing the person’s ability to live independently and having a significant impact on both the affected individual and society, as noted in [2]. Ms. Ashwini. S. Shindea and Dr. Veena. V. Desaan discusses and implements various machine learning algorithms, including Naive Bayes, Logistic Regression, Multi-layer Perceptron, Support Vector Machine, and Decision Tree, to analyse brain tumours using MRI images [3]. In [4], authors used structural magnetic resonance imaging (sMRI) 1) to increase the accuracy level that is comparable to the state-of-the-art methods; 2) to overcome the overfitting problem, and 3) to analyse proven landmarks of the brain that provide discernible features for AD diagnosis. They focused specifically on both the left and right hippocampus areas of the human brain and then applied simple convolutional neural networks (CNNs) as feature extractors and soft-max cross-entropy as the classifier. In [5], the author provides a brief review of some critical literature on AD and explores how deep learning can aid researchers in diagnosing the disease at its early stages. The data used from international challenge for automated prediction of MCI from MRI data to address the multiclass classification problem and a novel multiclass classification approach that addresses the outlier detection problem, uses pairwise t-test feature selection, project the selected features onto a Partial-Least-Squares multiclass subspace, and applies one-versus-one error correction output codes classification for detection of Alzheimer’s Disease in its early stages is proposed in [6]. In the author studies, specific information in clinical text, including the age, sex and genes of the person and apolipoprotein E; the brain function is established using resting-state functional data (MRI) for the measurement of connectivity in the brain regions [7]. A specialised network of autoencoders is used in earlier diagnosis to distinguish between natural ageing and disorder progression. The proposed approach effectively incorporates biased neural network functionality, enabling reliable recognition of Alzheimer’s disease. In [7], a model is proposed that takes brain MRI sample images as input and determines whether a person has mild, moderate, or no Alzheimer's disease as output [8]. Vandana B.S. and Sathyavathi R. Alva proposed two methods. Radiological feature extraction using image processing and machine learning from MRI images and Analysis of Alzheimer’s disease state by using a deep learning approach [9]. A deep learning-based method to extract valuable AD biomarkers from structural magnetic resonance imaging (sMRI) and classify brain images into AD, mild cognitive impairment (MCI), and cognitively normal (CN) groups. Adapted and trained convolutional neural networks (CNNs) on sMRI images of the brain from ADNI datasets available in online databases and proposed mechanism was used to combine features from different layers to hierarchically transform the images from magnetic resonance imaging into more compact high-level features and reduced number of parameters which reduces the computation complexity develop in [10]. In the authors develop a deep learning-based pipeline for the accurate diagnosis and stratification of Alzheimer's disease (AD) stages [11]. The analysis pipeline utilises a shallow Convolutional Neural Network (CNN) architecture and 2D T1-weighted Magnetic Resonance (MR) brain images to provide both global classification (i.e., standard vs. Mild Cognitive Impairment (MCI) vs. Alzheimer's disease (AD)) and local classification. The latter deals with an even more challenging task to stratify MCI into a Very Mild Dementia (VMD), mild dementia (MD), and Moderate Dementia (MoD) as the prodromal AD stage. III. METHODOLOGY In this section, three algorithms are introduced that are used to extract features from MRI images and facilitate their classification. The first and second algorithms are primarily used in object classification under image processing and computer vision. The third one is applicable in both agricultural and medical image classification, when very little difference is found among the classes. A. Algorithm in determining feature table i. Read an image and apply pre-process (denoise and resize) on it. ii. Apply the SURF algorithm on the image and extract the value and location of the five most prominent (five consecutive points of highest strength) feature points. iii. From the points (xi, yi), i =1, 2, 3, 4, 5; select the point (xh, yh) of maximum scale. iv. Determine the Euclidian distance between (xh, yh) and (xi, yi) as: , i ≠ m. v. Sort the Euclidean distances, di and put them in a vector, DSURF = [d1 d2 d3 d4]. vi. Create a table of vector DSURF for four types of images. vii. Apply steps 1 to 6 on using ‘corners feature extraction using FAST algorithm’. viii. Apply steps 1 to 6 on using ‘corners feature extraction using Harris–Stephen’s algorithm’. ix. Apply steps 1 to 6 on using ‘corners feature extraction using minimum eigenvalue algorithm’. x. Apply FIS, MLR, Naïve Bayes, SVM and FCM on the complete table. B. Algorithm in Determining Coherence of Two Images using DWT Read the image I of class-1 Select the dimension of the image as N International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249-8958 (Online), Volume-15 Issue-1, October 2025 20 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.E467314050625 DOI: 10.35940/ijeat.E4673.15011025 Journal Website: www.ijeat.org Resize the image as: N×N for i = 1: N, Si = ith row of the image for i=1 to N while length(y)>1 y= dwt (Si); % continue the DWT until the length of y is 1. end%end of while loop r(i) = y; % ith element/row of the column vector end % end of for loop while length(r) > 64 r = dwt(r); % continue the DWT until the length of r is 64 end v1 = r; % vector with the lowest frequency of image I of class-1 determine v2 the vector with the lowest frequency of image I of class-2 C12 = coherence (v1, v2); % The coherence of the image of class-1 and class-2 C. MLs in image Classification The entire operation of the training or test phase of ‘classification of Alzheimer MRI image’ used in this paper is shown in Fig. 1. [Fig.1: Entire Operation of Image Classification (Training or Test)] Two deep learning models, CNN and its variant VGG-19, are also applied for classifying Alzheimer's images to achieve better results. However, image processing and machine learning algorithms are tried first to obtain reasonable results with less processing time. The image data set of the paper is collected from Kaggle’s open and free dataset under the project: ‘Alzheimer_MRI_Detection_Densenet_PyTorch’ with URL: https://www.kaggle.com/code/alibidaran/alzhaimer-mri-dete ction-densenet-pytorch/input. IV. RESULT AND DISCUSSIONS First, the four strongest features are taken under four algorithms: (i) SURF of [12], (ii) Corners Features using FAST algorithm, (iii) Corners Features using Harris–Stephen’s algorithm, and (iv) Corners Features using minimum eigenvalue algorithm of [13]. For each category, 400 records are extracted, and among them, 350 are taken for training, and the remaining 50 records are used for validation. Few records are shown in table I, II, III and IV against four algorithms. TableI: SURF Records SURF Category SURF1 SURF2 SURF3 SURF4 29.04 57.83 61.84 95.38 Mild Impairment 61.16 118.89 135.27 154.93 Moderate Impairment 48.81 59.01 108.27 156.10 Very Mild Impairment 0.96 53.89 98.21 124.17 No Impairment 29.59 30.29 77.21 109.55 Mild Impairment 8.05 27.75 65.79 118.56 Moderate Impairment 22.42 61.09 134.51 154.05 Very Mild Impairment TableII: Records of Corners Features using FAST Algorithm Corners Features using FAST algorithm Category FAST1 FAST2 FAST3 FAST4 40.70 67.20 70.03 89.14 Mild Impairment 68.15 75.01 87.59 111.00 Moderate Impairment 10.29 145.60 180.23 197.64 Very Mild Impairment 20.24 66.18 86.35 96.67 No Impairment 2.83 35.01 42.30 216.02 Mild Impairment 41.18 53.14 106.71 171.64 Moderate Impairment 34.13 74.25 139.85 142.41 Very Mild Impairment Table-III: Records of Corners Features using Harris–Stephen’s Algorithm Corners Features using Harris–Stephen’s algorithm Category HS1 HS2 HS3 HS4 75.42 107.25 120.64 125.90 Mild Impairment 4.50 150.80 165.42 186.43 Moderate Impairment 52.03 143.45 144.84 148.47 Very Mild Impairment 34.06 97.00 97.89 100.88 No Impairment 58.35 198.24 205.65 211.68 Mild Impairment 74.28 76.08 90.19 116.11 Moderate Impairment 55.44 59.77 72.94 155.65 Very Mild Impairment TableIV: Records of Corners Features using Minimum Eigenvalue Algorithm Corners Features using minimum eigenvalue algorithm Category EIG1 EIG2 EIG3 EIG4 9.51 65.76 69.70 90.46 Mild Impairment 21.61 110.24 147.67 150.36 Moderate Impairment 51.41 93.50 194.35 198.57 Very Mild Impairment 70.38 72.81 87.19 101.66 No Impairment 58.13 197.93 205.68 210.93 Mild Impairment 36.88 45.18 72.98 100.17 Moderate Impairment 55.40 59.66 73.11 155.89 Very Mild Impairment The ‘parallel coordinate plots’ of the data points are shown in Fig. 2 (a) and (b) for 8 and 16 features, respectively. Such a plot is primarily used to visualise higher-dimensional (usually more than three-dimensional) data points. Enormous overlap is found in the plot, and the points are also highly uncorrelated. The phenomenon is also visualised from the scattered plot of Figure. 3(a) and (b), where only two variables are considered. The AUC from the ROC of the four classes of images is found to be less than 60%, as shown in fig. 4. It will be difficult to segregate pictures from a Single ML. Under these adverse conditions, we will apply different approaches to classify four types of images using feature data points (or records). Alzheimer Classifications Combining Machine Learning and Signal Processing 21 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.E467314050625 DOI: 10.35940/ijeat.E4673.15011025 Journal Website: www.ijeat.org (a)Eight Feature Points (b) Sixteen feature points (c) Legends [Fig.2: Parallel Plot of Feature Points] (a) SURF1 vs. FAST1 (b) HS1 vs. EIG1 [Fig.3: Scattered Plot of Data Points for Two Cases] [Fig.4: ROC of Four Categories of Data] The next part of the result deals with the Euclidean distance between feature points on the image. First, the five strongest SURF feature points are selected on each of the four types of images shown in Fig. 5(a). The Euclidean distance between ‘the point of lowest value’ and the other four points is evaluated, then sorted. The four sorted distances are shown in the bar graph of Fig. 5(b). Similar images and graphs are shown for ‘Corners Features points using FAST algorithm’ in the figure. 5(c)-(d), image and graph for ‘Corners Features points using Harris–Stephen’s algorithm’ case are shown in Fig. 5(e)-(f), and those of ‘Corners Features points using minimum eigenvalue algorithm’ are shown in Fig. 5(g)-(h). (a) SURF Feature Points (b) Bar of Euclidian distances (c) Corners Features Points using FAST Algorithm International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249-8958 (Online), Volume-15 Issue-1, October 2025 22 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.E467314050625 DOI: 10.35940/ijeat.E4673.15011025 Journal Website: www.ijeat.org (d) Bar of Euclidian distances (e) Corners Features Points using Harris–Stephens Algorithm (f) Bar of Euclidean Distances (g) Corners Features Points using Minimum Eigenvalue Algorithm (h) Bar of Euclidean Distances Fig.5: Features of Four Types of Images The MLR is applied to the entire data set of four classes, and the result is shown in Fig. 6. Since the amplitude has very low variation, the data set has Very low correlation, and MLR reveals abysmal performance. Next, we applied FIS (Mamdani rule with centroid method of defuzzification) on the previous dataset. The results of the four-test data are shown in the figure. 7(a)-(d) against four class. The performance is deemed poor by the FIS. Finally, FCM (fuzzy c-means clustering) is applied to the complete data set. The corresponding result is shown in Fig. 8(a), where the data points of the four classes are distributed randomly, and the centres of the classes are very close; hence, FCM also fails to provide good performance. The cost function of FCM becomes fixed at a very high value, as shown in Fig. 8(b). [Fig.6: Results of MLR] Alzheimer Classifications Combining Machine Learning and Signal Processing 23 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.E467314050625 DOI: 10.35940/ijeat.E4673.15011025 Journal Website: www.ijeat.org (a) Input Vector, v = [15.54;37.64;93.89;102.4;2.24;113.4;179.3;206.4;52.92; 90.37;112.6;150] for Class-1 (b) Input Vector, v = [47;56.88;72.92;102.4;50.08;66.03;120.6;159.7;53.84; 59.32;100.2;150] for Class-2 (c) Input Vector, v = [30.06;53.38;103.6;136.7;32.2;92.63;74.97;151;26.5;61.5; 92.67;123] for Class-3 (d) Input Vector, v = [34.9;63.88;114.9;94.82;42.72;66.06;107.6;132;31.97; 103.7;115.1;144.2] for Class-4 [Fig.7: Results of FIS From the Table of Four Attributes] (a) Data Points of FCM (c) Profile of the Objective Function Fig.8: Performance of FCM on Classification of Four Classes] Next, a one-dimensional DWT is applied to each row of the image until it is reduced to a single point. The entire image becomes a column vector, and then a one-dimensional DWT is applied to the column vector to reduce its size to 1×64. The resultant column contains the few lowest and prominent frequency components of the image. First, the idea is applied to ‘Moderate Impairment’ and ‘no Impairment’ images, then wavelet coherence is applied to them to find the phase difference between them. The corresponding results are shown in fig.9 for three cases. Very little difference in phase is found only at medium and high frequency. The profile of the discrete plot also has a minimal difference. Similar results are shown for the case of ‘Mild Impairment’ and ‘Very Mild Impairment’, as shown in Fig. 10. The expected difference on the coherence graph is also found to be very poor for this case. Such a slight difference is not applicable in image classification; however, the above concept is widely used in many object classification applications under image processing, as discussed in [14]. International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249-8958 (Online), Volume-15 Issue-1, October 2025 24 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.E467314050625 DOI: 10.35940/ijeat.E4673.15011025 Journal Website: www.ijeat.org (a) DWT Components (Experiment-1) (b) DWT Coherence (Experiment-1) (c) DWT Components (Experiment-2) (d) DWT Coherence (Experiment-2) (e) DWT Components (Experiment-3) (f) DWT Coherence (Experiment-3) [Fig.9: DWT Components and Their Coherence] (a) DWT Components (Experiment-1) (b) DWT Coherence (Experiment-1) Alzheimer Classifications Combining Machine Learning and Signal Processing 25 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.E467314050625 DOI: 10.35940/ijeat.E4673.15011025 Journal Website: www.ijeat.org (c) DWT Components (Experiment-1) (d) DWT Coherence (Experiment-1) (e) DWT Components (Experiment-1) (f) DWT Coherence (Experiment-1) [Fig.10: DWT Components and Their Coherence] Next, 100 images per class are used in BoF, with 80% allocated for training and 20% for testing. The program is run three times, and the corresponding frequency of visual code words is shown in Fig. 11(a)-(c). The heat map of the bag of features against three experiments is shown in Fig. 12 (a)-(c). The detection of 'No Impairment' is found above 95%, and that of 'Moderate Impairment' is found near 85%. The 'Mild Impairment' showed poor results, i.e. near 60%, and the case of 'Very Mild Impairment' showed inferior results, near 10%. The reason behind the case is that both the 'Mild Impairment' and 'Very Mild Impairment' cases are very close to the case of 'No Impairment', but the 'Moderate Impairment' image has a wide variation from the image of the 'No Impairment' case. Finally, the performance of BoF is measured by taking 2560 images per class (80% training and 20% test). (a) Experiment-1 (b) Experiment-2 (c) Experiment-3 [Fig.11: Visual Code Words of Bag of Feature Taking 4×100 Images] International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249-8958 (Online), Volume-15 Issue-1, October 2025 26 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.E467314050625 DOI: 10.35940/ijeat.E4673.15011025 Journal Website: www.ijeat.org (a) Experiment-1 (b) Experiment-2 (c) Experiment-3 [Fig.12: Heat-Map of Bag of Features Taking 4×100 Images] The BoF code extracted 8388608 features from 8192 images, where the K-Means clustering was used to create a 500-word visual vocabulary, and the number of clusters was set to k = 500. The average accuracy is now improved to 81%, and the inferior component ‘'Very Mild Impairment’ also improved considerably, as shown in fig. 13. (a) Visual code words of a bag of feature (b) Heat-map of a bag of Feature [Fig.13: Performance of BoF Taking 4×2560 Images] The four types of images, each with a size of 2739, 2572, 3200, and 3008, as shown in the bar graph of Fig. 14, are used for deep learning (DL), where 80% of the images are used for training and 20% for testing under PyTorch CNN and Keras VGG. Taking size of image 224*224, learning rate = 0.001, number of epochs = 10, batch size = 32, number of features=32 and kernel size=3 at both convolution layer we got the accuracy of 95.7% for PyTorch CNN and 97.1% for keras VGG. The confusion matrices of the test images are shown in Fig. 15(a) and (b). [Fig.14: Number of Images of Four Categories]