Inter national J our nal of Electrical and Computer Engineering (IJECE) V ol. 15, No. 2, April 2025, pp. 1557 ∼ 1571 ISSN: 2088-8708, DOI: 10.11591/ijece.v15i2.pp1557-1571 ❒ 1557 Enhanced automated Alzheimer’ s disease detection fr om MRI images by exploring handcrafted and transfer lear ning featur e extraction methods T ouati Menad 1 , Mohamed Bentoumi 1 , Ar ezki Larbi 1 , Malika Mimi 1 , Abdelmalik T aleb Ahmed 2 1 Department of Electrical Engineering and Laboratory of Signals and Systems, F aculty of Sciences and T echnology , Uni v ersity Abdelhamid Ibn Badis of Mostag anem, Mostag anem, Algeria 2 Institute of Electronics, Microelectronics and Nanotechnology (IEMN), Uni v ersit ´ e Polytechnique Hauts de France, Uni v ersit ´ e de Lille, Centre National de la Recherche Scientique (CNRS), V alenciennes, France Article Inf o Article history: Recei v ed Feb 1, 2024 Re vised Oct 10, 2024 Accepted No v 20, 2024 K eyw ords: Alzheimer’ s disease Classication Con v olutional neural netw ork Handcrafted features Machine learning Magnetic resonance imaging images T ransfer learning ABSTRA CT The rising pre v alence of Alzheimer’ s dis ease (AD) poses a signicant global health challenge. Early detection of AD enables appropriate and timely treat- ment to slo w disease progression. In this paper , we propose an enhanced proce- dure for automated AD detection from magnetic resonance imaging (MRI) im- ages, focusing on tw o primary tasks: feature e xtraction and classication. F or feature e xtraction, we ha v e in v estig ated tw o cate gories of methods: handcrafted techniques and those based on pre-trained con v olutional neural netw ork (CNN) models. Handcrafted methods are preceded by a preprocessing step to impro v e the MRI image contrast, while the pre-trained CNN models were adapted by utilizing only a part of the models as feature e xtractors, incorporating a global a v erage pooling (GAP) layer to atten the featur e v ector and reduce its dimen- sionality . F or classication, we emplo yed three dif ferent algorithms as binary classiers to detect AD from MRI images. Our results demonstrate that the support v ector machine (SVM) classier achie v es a classication accurac y of 99 . 92% with Gabor features and 100% with ResNet101 CNN features, compet- ing with e xisting methods. T his study underscores the ef fecti v eness of feature e xtraction using Gabor lters , as well as those based on the adapted pre-trained CNN models, for accurat e AD detection from MRI images, of fe ring signicant adv ancements in early diagnosis. This is an open access article under the CC BY -SA license . Corresponding A uthor: T ouati Menad Department of Electrical Engineering and Laboratory of Signals and Systems, F aculty of Sciences and T ech- nology , Uni v ersity Abdelhamid Ibn Badis of Mostag anem Route Belahcene.Bp 277, Mostag anem, Algeria Email: touati.menad.etu@uni v-mosta.dz 1. INTR ODUCTION Alzheimer’ s disease (AD) is a progressi v e neurode generati v e condition that primarily impacts the brain, resulting in a gradual deterioration of memory , cogniti v e abilities and social aptitude. From a structural perspecti v e of the brain, AD is characterized by brai n shrinkage and e v entual neuronal death, rendering it the foremost cause of dementia [1]. AD represents a distinct and pathological condition be yond what is considered normal aging, yet the lik elihood of de v eloping AD rises as indi viduals gro w older . Approximately 5% of indi viduals aged 65 to 74 years are af fected by AD, while nearly 50% of those aged 85 and older suf fer from J ournal homepage: http://ijece .iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
1558 ❒ ISSN: 2088-8708 the disease [2]. In 2020, it w as estimated that around 50 million people w orldwide were li ving with AD [3], [4]. This number is e xpected to reach approximately 131.5 million people w orldwide by 2050 [5]. The early detection of AD represents one of the most intricate challenges for neurologists. V ari- ous brain imaging modalities, including magnetic resonance imaging (MRI), computed tomograph y (CT), and positron emission tomograph y (PET), allo w for the identication of structural and functional changes associ- ated with AD. Ho we v er , the manual e xamination of images by doctors or radiologists is often time-consuming and susceptible to errors. The de v elopment of automated diagnostic aid systems pro vides v aluable support to healthcare professionals, f acilitating the early detection of AD and enabling quick er and more accurate diag- noses while reducing medical errors and enhancing treatment outcomes. According to the literature, AD detection methods are based either on a single imaging modality or on multimodal approaches, particularly combining MRI with PET . Multimodal techniques can be cate gorized into tw o types: those that fuse features e xtracted from both im aging modalities [6], [7] and those that mer ge MRI and PET images [8]–[10]. The latter approach, while requiring highly comple x image fusion techniques, is more ef fecti v e for tracking the progression of AD. Ho we v er , PET , as an in v asi v e modality in v olving a radioac- ti v e tracer [ 6], [7], is often less f a v ored com p a red to MRI alone in the conte xt of AD detection. MRI is the most widely used imaging modality [11] due to its non-in v asi v e nature and its capacity to pro vide high-resolution structural information about the brain. In the conte xt of AD detection from MRI im ages, the process encompasses three k e y stages: image preprocessing, feature e xtraction, and classication. The preprocessing steps may include denoising, contrast enhancement, and/or se gmentation to detect and localize the re gion of interest (R OI). Se gmentation is particu- larly benecial for the detection and identication of brain tumors [12]. Ho we v er , in the conte xt of Alzheimer’ s disease detection, se gmentation is not strictly necessary , as AD af fects the entire brain. Nonetheless, it becomes rele v ant when applying methods for e xtracting morphological features from the brain [13]. Feature e xtraction is a transformation operation that con v erts an image (2D) into a feature v ector (1D) that represents its informa- tion. Feature e xtraction methods are generally classied into handcrafted methods and CNN-based methods. The classication step assigns observ ations to predened cate gories or classes based on their feature v ectors. In this paper , we present an enhanced procedure for automated AD detection from MR I images. Our approach comprises tw o primary steps: feature e xtraction and classication. F or feature e xtraction, we in v esti- g ate se v eral methods: three handcrafted methods (his togram of oriented gradients (HOG), local binary patterns (LBP) and Gabor lt ers) and nine pre-trained CNN models [14] (V GG16, Ale xNet, ResNet101, GoogLeNet, DenseNet, InceptionV3, SqueezeNet, MobileNetV2 and Shuf eNet). Handcrafted methods are preceded by a ltering-based pre-processing step to impro v e MRI image quality before applying the e xtractors. The pre- trained CNN models are adapted by adding a gl obal a v erage pooling (GAP) layer without ne-tuning the net- w ork parameters. F or the classication step, we emplo yed three classiers: support v ector machine (SVM), k- nearest neighbors (KNN) and decision trees (DT). These classiers are used to distinguish between Alzheimer’ s disease (AD) and normal cases (cogniti v ely normal, CN) classes from MRI images. Our results are compared with those presented in related w orks. Our major contrib utions in this paper are summarized as − W e e xplored tw o approaches for feature e xtraction from MRI images: the handcrafted approach and the transfer learning (TL) approach. − W e utilized three classiers—SVM, KNN, and DT—to classify AD and CN subjects, enabling us to identify the optimal combination of feature e xtractor and classier . − W e used three publicly a v ailable databases containing MRI images via the Kaggle platform. − T o assess the generalization ability of each e xtractor -classier combination, we applied k-fold cross-v alidation. The remainder of the paper is or g anized as follo ws. Section 2 re vie ws related w ork in the eld of AD detection from MRI images. Section 3 details the proposed methodology , outlining the v arious steps in v olv ed. In section 4, we present the e xperimental setup and results, follo wed by a discussion comparing our ndings with state-of-the-art methods. Finally , section 5 concludes the paper . 2. RELA TED W ORK Automatic detection and diagnosis of Alzheimer’ s disease (AD) are major challenges in the eld of neural me dical research. In this conte xt, se v eral researchers ha v e presented v arious models and approaches for the autom atic detection and diagnosis of AD from MRI images. Li and Y ang [15] used MRI images from Int J Elec & Comp Eng, V ol. 15, No. 2, April 2025: 1557-1571 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 1559 the Alzheimer’ s disease neuroimaging initiati v e (ADNI) database for tw o types of subjects, AD and CN. Three machine learning-based classiers were emplo yed to predict Alzheimer’ s disease and identify the re gions of the brain af fected by this disease. A comparison study w as conducted among the three distinct classiers: SVM, V GGNet and ResNet. The accurac y v alues for the AD to CN data classication across the three classiers were: support v ector machine ( 90% ), V GGNet ( 95% ), and ResNet ( 95% ). Zhang et al. [16] e xtracted tw o types of features from MRI images: gray matter (GM) v olume and lateral ization inde x (LI), using h ypothesis testing. The study included four data classes from the ADNI database: CN, early mild cogniti v e impairment (EMCI), late mild cogniti v e impairment (LMCI) and AD. Subsequently , se v eral classication algorithms were emplo yed, including random forest (RF), decision tree (DT), k-nearest neighbor (KNN) and support v ector ma- chine (SVM) with linear , RBF and polynomial k ernel. F or tw o groups of subjects—AD group v ersus CN—the SVM classier with a linear k ernel and the KNN classier achie v ed the highest accuracies of 98 , 09% and 98 , 25% , respecti v ely . Araf a et al. [17] applied deep learning (DL) to detect and diagnose AD using tw o con v olutional neural netw ork models: a custom end-to-end CNN de v eloped from scratch and a ne-tuned V GG16 model. The implementation in v olv ed three stages : dataset preparation with image size reduction, data augmentation, and model training/testing with an 80% / 20% split. Ev aluation on a subset of MRI images from the ADNI database re v ealed that the custom CNN achie v ed an accurac y of 99 . 95% , while the V GG16 model attained 97 . 44% . Naz et al. [18] emplo yed machine learning (ML) and deep learning to detect and identify Alzheimer’ s disease. The y proposed a system of CNN-based architectures using features e xtracted from MRI images of the entire ADNI database, which contains three dif ferent class types (AD, CN and mild cogniti v e impairment (MCI)). The classication w as performed out by the SVM classier on the three classes distrib uted as follo ws: AD/MCI, CN/MCI, and AD/CN. The results reached an accurac y of 99 . 27% (MCI/AD), 98 . 89% (AD/CN), and 97 . 06% (MCI/CN). The CNN-based approach w as also utilised by Y ousry AbdulAzeem et al. [19] on the ADNI database containing tw o-class MRI images AD and CN. A data augmentation technique w as emplo yed to increase the number of data. Feature e xtraction and classication were performed using an end-to-end CNN, with cross-v alidation allocating 85% of the data for training, 10% for v alidation, and 5% for testing. The achie v ed classication accurac y for AD/CN w as 97 . 80% . Ismail et al. [20] impl emented a mul timodal image fusion method to mer ge MRI images with a modular set of image pre-processing procedures. This method w as applied to the ADNI database, which includes tw o classes: AD and CN. T o e xtract rele v ant and generic information from the fused images, a 3D CNN netw ork w as utilized. The characteristics of both classes were classied using three classiers: CNN, SVM and RF . The AD/CN classication yielded accuarc y v alues of 98 . 21% , 91% , and 85 . 90% , respecti v ely . Rang araju et al. proposed in their research paper [21] an end-to-end CNN model for the automatic identication of Alzheimer’ s disease using 3D brain MRI data. The model comprises three main components: First, a patch con v olutional neural netw ork (PCNN) is emplo yed to e xtract discriminati v e features from each MRI patch. Second, an octa v e con v olution layer is utilized to reduce spatial redundanc y and e xpand the recept i v e eld for capturing detailed brain structure. Finally , a dual attention-a w are con v olutional classier further renes the feature representation to enhance the accurac y of AD detection. It is w orth noting that the MRI data is pre-processed, which includes image scaling and denoising. The designed end-to-end CNN model achie v ed a test accurac y of 99 . 87% for cate gorizing deme n t ia stages using the publicly a v ailable Alzheimer’ s disease neuroimaging initiati v e (ADNI) dataset. Referring to T abl e 1, it is e vident that there are still impro v ements to be made in the automation proce ss for detecting Alzheimer’ s disease from MRI images. In pre vious w orks based on DL methods, con v olutional neural netw ork models ha v e often been trained or ne-tuned on small image databases. Ho we v er , con v olutional neural netw orks require lar ge datasets (big data) for ef fecti v e learning. T o address this limitation, we propose to use pre-trained CNN models as feature e xtractors without readjusting the netw ork parameters. Moreo v er , our w ork distinguishes itself through the application of Gabor lters as feature e xtractors on MRI images for AD detection. T o our kno wledge, this is the rst study to in v estig ate this approach, thereby opening ne w perspecti v es in the eld of biomedical image analysis. Furthermore, another signicant moti v ation behind our w ork is to propose a straightforw ard and accessible procedure that circumv ents the use of comple x methods, such as those based on multimodal or 3D MRI analysis. By simplifying the tools emplo yed, we ai m to enhance detection ef cienc y while ensuring greater applicability in clinical settings where resources and time are often constrained. Enhanced automated Alzheimer’ s disease detection fr om MRI ima g es by ... (T ouati Menad) Evaluation Warning : The document was created with Spire.PDF for Python.
1560 ❒ ISSN: 2088-8708 T able 1. Summary of the state-of-the-art for AD detection Authors and Feature e xtractor Database Classication Cross-v alidation Accurac y (%) references method Li and Y ang CNN AD-CN (ADNI) SVM(TL) 85% training and 90 (2021) [15] dataset 3D-V GGNet (end to end) 15% test 95 3D-ResNet (end to end) 95 Zhang et al. GM AD-CN-MCI (ADNI) SVM 10-fold 98.09 (2022) [16] LI dataset RF 94.60 GM+LI DT 91.10 KNN 98.25 Araf a et al. CNN AD-CN (ADNI) CNN (end to end) 80% training and 99.95 (2023) [17] dataset V GG16 (end to end) 20% test 97.44 Naz. et al. Ale xNet(con v5) AD-CN-MCI SVM 80% training, 91.38 (2021) [18] V GG16(FC6) (ADNI) dataset KNN 10% v alidation 98.89 V GG19(FC6) and 10% test 99.27 AbdulAzeem CNN AD-CN (ADNI) CNN (end to end) 95% training 97.80 et al. dataset and 5% test (2021) [19] Ismail et al. CNN AD-CN (ADNI) SVM 10-fold 91.00 (2022) [20] dataset RF 85.90 3D CNN (end to end) 98.21 Rang araju et al. CNN EMCI-LMCI- 3D-CNN (end to end) Holdout 99.87 (2024) [21] MCI-AD-CN (ADNI) dataset 3. DESCRIPTION OF THE METHODOLOGY In this paper , we propose an enhanced automated procedure for AD detection using machine learning techniques applied to an MRI image dataset, aiming to achie v e high-performance results in the AD detection. The proposed procedure can be di vided into t w o k e y phases: feature e xtraction and cl assication. W e e xplored tw o approaches for feature e xtraction from MRI images: the handcrafted approach and the transfer learning (TL) approach [22]. As il lustrated in Figure 1, the procedure is di vided into tw o distinct pipelines, each corresponding to the implementation of one of the considered approaches for feature e xtraction. The handcrafted approach in v olv es a tw o-step process. First, the input image is ltered to impro v e its contrast. Subsequently , feature e xtraction transforms the ltered images into feature v ectors. W e e xplored three handcrafted methods: HOG, LBP , and Gabor methods. On the other hand, the transfer learning approach uses pre-trained CNN models as feature e xtractors. Finally , a classication process is performed on the feature v ectors using three dif ferent classi ers: SVM, KNN, and decision tree (DT). In the follo wing paragraphs, a brief o v ervie w of all these methods is pro vided, preceded by a short description of the MRI image datasets used in this w ork. Figure 1. Block diagram of the proposed frame w ork 3.1. Description of the databases The rst crucial stage in a machine learning process is data collection. The quality of the training data is essential to ensure the accurac y of predictions made by machine learning systems. In this w ork, we ha v e used Int J Elec & Comp Eng, V ol. 15, No. 2, April 2025: 1557-1571 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 1561 three publicly a v ailable databases from the Kaggle platform, each one with an unbalanced distrib ution of MRI images. The rst database contains 5121 images of 176 × 208 pix els, di vided into four classes: NonDemented, MildDemented, Moder ateDemented, and V eryMildDemented . The second database consists of 6163 images of dimensions 369 × 369 × 3 di vided into three cate gories: NonDemented, MildDemented, and V eryMildDemented . The third database, kno wn as the Alzheimer’ s disease neuroimaging initiati v e (ADNI) entails 5154 images of v arying sizes ( 170 × 256 , 166 × 256 , and 160 × 260 pix els) and comprises three classes: Alzheimer’ s disease (AD), mild cogniti v e impairment (MCI) and cogniti v ely normal (CN). T o e v aluate our proposed method, we ha v e combined some cate gories from the three databases into tw o distinct labels, Alzheimer’ s disease (AD) and normal cases (cogniti v ely normal, CN) for AD detection. The selection gi v en includes 400 images from each database di vided into 200 AD images and 200 CN images. Thus, the nal dataset consists of 1200 grayscale 2D MRI images with tw o distinct cate gories: 600 AD images and 600 CN ones. All images are resized to ( 256 × 256 ) pix els to ensure size uniformity , as reported in T able 2. T able 2. Distrib ution of selected MRI images from the three databases. Database and format Labeling # of images Size of images Database 1 (MRI images) format JPEG Cogniti v ely Normal (CN) 200 256 × 256 Alzheimer’ s disease (AD) 200 Database 2 (MRI images) format JPEG Cogniti v ely Normal (CN) 200 256 × 256 Alzheimer’ s disease (AD) 200 Database 3 (MRI images) ADNI format PNG Cogniti v ely Normal (CN) 200 256 × 256 Alzheimer’ s disease (AD) 200 3.2. F eatur e extraction The feature e xtraction phase in an automated Alzheimer’ s disease (AD) detection procedure using MRI images is crucial, as the quality of the features directly impacts the performance of the process. This phase can be considered as a transformation process from a 2D ima g e to a 1D v ector , where each element of the v ector represents a rele v ant feature of the image. In proposed w ork, we ha v e e xplored se v eral feature e xtraction methods belonging to tw o cate gories of approaches, as pre viously mentioned. This represents the rst suggested main contrib ution. The follo wing paragraphs present the methods used in this phase. 3.2.1. Handcrafted extractors The handcrafted methods are applied follo wing a pre-processing step that in v olv es ltering. This step is crucial for impro ving the initial quality of MRI images by reducing noise and enhancing contrast which allo ws the handcrafted feature e xtraction methods to be more ef fecti v e. In our w ork, we ha v e opted to use a median lter [23], [24] on the MRI images due to its balance of simplicity and ef fecti v eness. The median lter is particularly well-suited for medical imaging as it ef fecti v ely remo v es nois e while preserving edges, which are vital for maintaining the inte grity of the image structures. The enhanced image quality directly contrib utes to the rob ustness and accurac y of the entire detection procedure based on handcrafted feature e xtraction methods. Subsequently , we ha v e applied handcrafted feature e xtraction methods. In this w ork, we ha v e utilized well- kno wn methods, namely the histogram of oriented gradients (HOG) [25], local binary patt erns (LPB) [26] [27], and Gabor lters [28], which are introduced ne xt. a. Histogram of oriented gradients The histogram of oriented gradients (HOG) [25] is a feature e xtraction operator used for object detec- tion in i mages. The HOG descriptor quanties and represents the te xtures and shapes present in an image. F or each pix el, the intensity gradient is calculated in both horizontal and v ertical directions, as sho wn by (1): G x = ∂ I ( x,y ) ∂ x G y = ∂ I ( x,y ) ∂ y (1) Where I is the image, G x is the gradient in the horizontal ( x ) direction and G y is the gradient in the v ertical ( y ) direction. In practice, these gradients can be approximated using con v olution lters such as the Sobel lter [29]. These gradients are then con v erted into magnitudes and orientations. The image is di vided into small cells, typically 8 × 8 pix els. F or each cell, a histogram of gradient orientations is constructed with the gradients weighted by their magnitudes, the cells are then grouped into blocks (e.g., 2 × 2 cells). The histograms of the cells within a block are normalized which helps to mak e the descriptor less sensiti v e to changes in lighting. Finally , the normalized histograms of all the blocks are concatenated to form a feature v ector representing the entire image. Enhanced automated Alzheimer’ s disease detection fr om MRI ima g es by ... (T ouati Menad) Evaluation Warning : The document was created with Spire.PDF for Python.
1562 ❒ ISSN: 2088-8708 b . Local binary patterns The LBP method is a technique for e xtracting te xture features from images [26], [27]. Its fundamental principle in v olv es comparing pix el intensities. F or each pix el in an image, the method compares the intensity of the pix el with that of its neighboring pix els, typically within a 3 × 3 neighborhood in Figure 2. If a neighbor’ s intensity is greater than or equal to the central pix el’ s intensity , a 1 is assigned; otherwise, a 0 is assigned. This binary comparison is performed for each neighbor , thereby forming a binary pattern around the central pix el. Figure 2. An e xample of calculating an LBP v alue These binary bits are then combined to form an 8-bit binary number in the case of a 3 × 3 pix el neigh- borhood. This number is con v erted into a decimal v alue, representing a unique LBP pattern. T o construct the image’ s feature v ector , a histogram of the occurrences of these decimal v alues is constructed. This histogram represents the te xture patterns present in the image and serv es as features of the original image. c. Gabor lters Gabor lters is a technique emplo yed as feature e xtraction method in image processing to e xtract te xtural and structural information from images. The proces s of designing a feature v ector using Gabor lters in v olv es se v eral k e y steps [28], [30]. First, Gabor lter s are constructed using sinus oidal functions modulated by a Gaussian function, as described by (2) and (3): G ( x, y ; σ , θ ) = exp − x 2 + y 2 2 σ 2 · cos 2 π x λ (2) with: x = m cos θ + n sin θ y = − m sin θ + n cos θ (3) where m and n are the coordinates of a pix el in the image with size ( M , N ) . The parameter σ controls the scale of the lter , while θ controls its orientati on. F or each combination of scale σ and orientation θ , we obtain a distinct Gabor lter . Ne xt, each Gabor lter is a p pl ied to the image, producing a ltered image, which results in a series of feature maps corresponding to each lter in Figure 3. The feature v ector can then be constructed in v arious w ays. One approach is to calculate global statistics, such as the mean or v ariance of the l ter responses. Alternati v ely , the responses can be directly concatenated, or histograms of the r esponses can be created to capture their di strib ution. Finally , the feature v ector is often normalized to ensure that the v alues are comparable and to minimize the ef fects of scale or lighting v ariations. This feature v ector is subsequently used for classication tasks. Figure 3. An e xample of a Gabor lter response with 5 scales and 8 orientations for an MRI image Int J Elec & Comp Eng, V ol. 15, No. 2, April 2025: 1557-1571 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 1563 3.2.2. CNN extractors F or the transfer learning method, we ha v e used nine distinct pre-trained CNN models: Ale xNet, ResNet101, DenseNet201, GoogLeNet, SqueezeNet, InceptionV3, V GG16, MobileNetV2, and Shuf eNet [31]. These models were pre-trained on the e xtensi v e ImageNet database [32] which comprises o v er 14 mil- lion images distrib uted across 1,000 dif ferent classes. Each CNN model consists of tw o main parts: a feature e xtraction part and a classication part as sho wn in Figure 4. Figure 4. An e xample of CNN netw ork architecture The feature e xtraction part is composed of a series of blocks. Each block includes con v olutional laye rs that e xtract hierarchical features from the input images, as well as pooling layers to reduce dimensionality . The feature maps produced by these blocks are then processed by a nonlinear acti v ation function, such as the rectied linear unit (ReLU) [12], [33]. The classication part consists of fully connected (FC) layers with the last FC layer emplo ying the Softmax function for classication. In our w ork, we ha v e utilized the nine pre-trained models mentioned earlier within the transfer le arning (TL) frame w ork to ef ciently e xtract feature v ectors from MRI images. This approach a v oids the need to design ne w CNN models from scratch which requires a la r ge database and signicant computational resources. W e ha v e carefully adapted these models to enhance performance and to reduce both the cost and training time required. Specically , based on prior w ork [33], we identied and remo v ed the classication component and the nal pooling layer from the feature e xtraction part of each pre-trained CNN model. As a result, at the output of the remaining feature e xtraction section, we obtained a set of feature maps { S i } with dimensions M and N , which depend on the characteristics of the l ast con v olutional layer for each pre-trained CNN model. Flattening this set of features to create a feature v ector results in a v ery high dimensionality . Therefore, to enhance classication performance, it w as necessary to reduce the dimensionali ty of the feature v ectors by incorporating a global a v erage pooling layer into the retai n e d part of each pre-trained CNN model. This operation attens and reduces the size of the feature v ectors in a single step as sho wn in Figure 5. The global a v erage pooling (GAP) is described as (4): x i = P M m P N n S i ( m, n ) M × N ; i ∈ [1 , p ] (4) Where ( M , N ) are the size of the last p feature maps { S i } from the retained part. These results in a feature v ector X of dimension p for each image, re g ardless of its size. This method does not require retraining or ne- tuning the pre-trained CNN models. The obtained feature v ectors are then used in the subsequent procedure for detecting Alzheimer’ s disease (AD) from MRI images. Figure 5. Pre-trained models used for feature e xtraction Enhanced automated Alzheimer’ s disease detection fr om MRI ima g es by ... (T ouati Menad) Evaluation Warning : The document was created with Spire.PDF for Python.
1564 ❒ ISSN: 2088-8708 3.3. Classication Detection of Al zheimer’ s disease from MRI images is re g arded as a binary classication problem (positi v e class: AD and ne g ati v e class: CN). Classication can be based on either supervised or unsupervised methods. When we ha v e labeled data (labeled observ ations), we are addressing a supervised classication problem, as is the case in this w ork. Con v ersely , if the data is unlabel ed, it w ould represent an unsupervised classication scenario. In our study , we emplo yed three supervised classication methods for binary classi- cation: support v ector machine (SVM) [34], [35], k-nearest neighbors (KNN) [36] and decision tree (DT) [37] which are briey introduced in the follo wing paragraphs. a. Support v ector machine (SVM) classier A support v ector machine (SVM) is a machine learning algorithm introduced by Vladimir V apnik [34]. SVM aims to nd an optimal linear h yperplane separating tw o classes. Its principle is based on maximizing the mar gin between the data d i strib utions of the tw o classes in the feature space (the distance between the tw o classes) while minimizing classication errors [38]. F or an SVM classier , we consider a training set D consisting of N e xamples ( X i , y i ) with X i ∈ R p belonging to one of the tw o classes and labeled by y i ∈ { +1 , − 1 } . The separating h yperplane H can be dened by equation (2), where w ∈ R p and b ∈ R represent the parameters of the separating h yperplane. H : ⟨ w , X i ⟩ + b (5) The v alues of w and b are determined through learning by minimizing the criterion J (eq. 3.3.) under the follo wing constraints: min w ,b J = 1 2 w 2 under the constraints: y i ( ⟨ w , X i ⟩ + b ) ⩾ 1 ; i = 1 · · · N (6) b . The k-nearest neighbor (KNN) classier The k-nearest neighbor (KNN) classier is a non-parametric supervised learning algorithm that clas- sies data based on the proximity of points in the feature space. Initially de v eloped by Ev elyn Fix and Joseph Hodges in 1951, and later e xtended by Thomas Co v er in 1967, KNN operates by identifying the k nearest neighbors of a data point to be classied, using a distance measure, often Euclidean distance. KNN then as- signs the majority class among these neighbors to the data point in question. The method consists of tw o main steps: determining the nearest neighbors and assigning the class based on these neighbors. c. Decision tree (DT) classier The decision tree is a supervised learning algorithm used for classication and is often applied to image feature v ectors. It b uilds models in a tree structure where each node represents a test on a feature of the input. The branches of the tree correspond to possible v alues of the attrib utes, while the lea v es denote the nal decisions or predicted classes. The decision tree recursi v ely partitions the data space based on e v aluation criteria to select the best splitting features, emplo ying heuristics to pre v ent o v ertting. This model ef fecti v ely classies data by constructing a series of tests based on numerical attri b ut es compared to predened thresholds. 4. EXPERIMENT A TION AND RESUL TS The objecti v e of this w ork is to de v elop an automated procedure for detecting Alzheimer’ s disease from MRI images. As pre viously mentioned, this procedure consists of tw o main phases: feature e xtraction and classication. T o desi gn and e v aluate it, we follo wed a series of steps outlined in the o wchart sho wn in Figure 6. The source code for this methodology w as de v eloped within the MA TLAB en vironment. First, we constructed an MRI image database by mer ging three distinct datasets, as detailed in section 3.1. These publicly a v ailable datasets from Kaggle dif fer in image quality , dimensions and classes. Images originally in color were con v erted to grayscale, and then resized to 256 × 256 pix els to ensure uniformity . W e randomly selected 400 images from each dataset, with 200 images per cate gory (AD and CN), ensuring that the nal mer ged dataset maintains high quality and is free from notable artif acts. This process resulted in a database comprising 1200 MRI images, e v enly di vided into tw o classes: 600 images in the AD class (positi v e class) and Int J Elec & Comp Eng, V ol. 15, No. 2, April 2025: 1557-1571 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 1565 600 images in the CN class (ne g ati v e class). W e then applied feature e xtraction methods to all the images in this database, utilizing three handcrafted methods as well as nine pre-trained CNN models, each serving as a feature e xtractor as sho wn in Figure 6. The handcrafted methods are preceded by a pre-processing step based on ltering to enhance the quality of the MRI image, while the pre-trained CNN models ha v e only been adapted by the addition of a global a v erage pooling (GAP) layer without an y ne-tuning netw ork parameters. Figure 6. Detailed synoptic of the dif ferent stages of the w ork process Ne xt, we e xamined all possible combinations of feature v ectors by pairing them with three classiers : SVM, KNN, and DT . This resulted in 36 dif ferent combinations, enabling a thorough e v aluation of the imple- mented procedure. T o assess the performa n c e of these combinations, and consequently the o v erall procedure for detecting AD from MRI images, we ha v e emplo yed se v eral metrics, which will be detailed in the follo wing subsection. 4.1. P erf ormance metrics and v alidation T o objecti v ely e v aluate the performance of our procedure, we use se v eral k e y metrics: accurac y (A CC), sensiti vity (SEN), and specicity (SPE). These metrics are calculated by comparing our predicted outputs with the actual data. The classication results are cate gorized into four types: true positi v e (TP), true ne g ati v e (TN), f alse positi v e (FP), and f alse ne g ati v e (FN) [38]. The metrics are dened as follo ws: − Accurac y (A CC) is calculated by the formula: A CC = T P + T N T P + T N + F P + F N (7) − Sensiti vity (SEN) is gi v en by: Sensiti vity = T P T P + F N (8) − Specicity (SPE) is computed using: Enhanced automated Alzheimer’ s disease detection fr om MRI ima g es by ... (T ouati Menad) Evaluation Warning : The document was created with Spire.PDF for Python.
1566 ❒ ISSN: 2088-8708 Specicity = T N T N + F P (9) These metrics are emplo yed to assess the performance of the SVM, KNN, and DT classiers. In this study , we estimate these performance metrics using the k-fold cross-v alidation method [33] [37]. The dataset is di vided into k subsets. W e train the model on k − 1 of these subsets while testing (i.e., e v aluating performance metrics) on the remaining subset. This process i s repeated k times, with each subset serving as the test set once. The global accurac y (GA) is the a v erage of the performance metrics obtained across all k iterations, calculated as (10): GA = 1 k k X i =1 A CC i (10) Similarly , the a v erage sensiti vity ( G sen ) and a v erage specicity ( G spe ) are computed across all k iterations, calculated as (11): G sen = 1 k k X i =1 SEN i (11) G spe = 1 k k X i =1 SPE i (12) W e apply ten-fold cross-v alidation ( k = 10) to e v aluate our proposed procedure. 4.2. Results and analysis In this subsection, we present the main results obtained through our procedure for detecting Alzheimer’ s disease from MRI images. It is important to emphasize that the primary objecti v e of this procedure is to distin- guish between images representing Alzheimer’ s disease (AD) and those of normal cases (cogniti v ely normal, CN). T able 3, along with Figures 7 to 9, pro vides a comparison of cl assication performance using v arious combinations of feature e xtractors and classiers. Among the handcrafted features used, Gabor features combined with the SVM classier achie v ed the best o v erall performance, with an accurac y (GA) of 99 . 92% , sensiti vity ( G sen ) of 99 . 83% , and specicity ( G spe ) of 100% . Although LBP and HOG features also e xhibited good performance, their results were slightly lo wer compared to the other features. LBP and HOG feature e xtraction methods are particularly ef fecti v e at capturing local te xture patterns; ho we v er , this may not be suf cient to address the comple xity of MRI images in the conte xt of AD detection. Among the pre-trained CNN models tested, ResNet101 demonstrated e xceptional performance, achie v- ing a 100% accurac y when combined with the SVM classier . Other models, such as DenseNet201, SqueezeNet, and Ale xNet, also e xhibited e xcellent performance, with accurac y rates e xceeding 99% in most cases , partic- ularly when paired with the SVM classier . In contrast, features e xtracted using the V GG16 model sho wed relati v ely weak er performance in comparison. Specically , the V GG16 model combined with the DT classier produced modest results, with global accurac y , global sensiti vity , and global specicity all rated at 91 . 33% . Re g arding the classiers, the SVM pro v ed to be the most ef fecti v e when combined with v arious feature e xtractors, including pre-trained CNN models, achie ving the highest scores in o v erall accurac y , sensiti vity , and specicity . The KNN classier also demonstrated solid performance, though it w as slightly less ef fecti v e than the SVM. Ho we v er , KNN outperformed SVM when used with Gabor and DenseNet201 feature e xtractors. In contrast, the DT classier sho wed more v ariable results, with accurac y rates sometimes f alling belo w 95% , making it less ef fecti v e compared to the SVM and KNN classiers. 4.3. Comparison with state-of-the-art methods In this subsection, we compare the performance of ou r enhanced procedure with that of related w orks for Alzheimer’ s disease detection. It is crucial to note that pro viding comparisons to other related w orks is challenging due to the dif fering protocols and image databases used for assessment. T o ensure a f air compar - ison, we focused on studies that closely align wit h our conte xt, specically those emplo ying a single imaging Int J Elec & Comp Eng, V ol. 15, No. 2, April 2025: 1557-1571 Evaluation Warning : The document was created with Spire.PDF for Python.