Alzheimer’s disease classification using attention mechanism and global average pooling on a convolutional neural network

• Main Author: Abd. Hamid, Nur Amirah
• Format: Thesis
• Language: English
• Published: 2022
• Subjects: T Technology (General)
• Online Access: http://eprints.utm.my/id/eprint/99688/1/NurAmirahAbdHamidMMJIIT2022.pdf; http://eprints.utm.my/id/eprint/99688/; http://dms.library.utm.my:8080/vital/access/manager/Repository/vital:150833

The robustness of Convolutional Neural Network (CNN) architecture as the innovative technology has led to the surge of research adoption for Alzheimer’s disease (AD) classification. CNN is replacing the conventional machine learning methods to assist and support experts in diagnosing AD. However, the performance of conventional CNN architecture in classifying AD class and Normal Control (NC) class is hindered by its behaviours that require a large-scale dataset. Nonetheless, the major hindrance in the AD domain is limited amount of dataset. Therefore, previous studies have adopted data-centric enhancement modules such as pre-processing techniques, data augmentation, and transfer learning strategies to improve the classification performance of CNN. Yet, these modules alone are still struggling to offer sound accuracy of classification of the disease due to CNN's overfitting issue and behaviour which is insensitivity to the local position information, also known as spatial invariance. A recent trend in this domain is the merge of an attention mechanism with CNN to enhance the classification performance. This is done by identifying and extracting the salient discriminative features of MRI images. However, the generalization ability is still hindered due to validity of one specific dataset found among many research works. This research then proposes a novel attention-based CNN model (AGap-CNN), that employs the global average pooling (GAP) to reduce the number of learning parameters to be used for classification by Softmax. The AGap-CNN combines the attention mechanism with the GAP layer for classification at the model header to enhance the classification performance of CNN and improve the generalization capability of the network. The AGap-CNN was validated on two benchmark datasets of OASIS and ADNI. Furthermore, in further analysing the network performance, the AGap-CNN was compared to the existing state-of-the-art methods. The proposed AGap-CNN model outperformed the existing state-of-the-art methods for the OASIS and ADNI datasets with 99.22% and 100% average validation accuracy, respectively. In other words, the proposed AGap-CNN model works with acceptable accuracy, sensitivity, and specificity in classifying AD class and NC class for both benchmark datasets of OASIS and ADNI dataset.