Deep learning model for prediction of progressive mild cognitive impairment to Alzheimer's disease using structural MRI / Lim Bing Yan

• Main Author: Lim, Bing Yan
• Format: Thesis
• Published: 2021
• Subjects: TJ Mechanical engineering and machinery
• Online Access: http://studentsrepo.um.edu.my/13016/1/Lim_Bing_Yan.jpg; http://studentsrepo.um.edu.my/13016/8/bing_yan.pdf; http://studentsrepo.um.edu.my/13016/

Alzheimer’s disease (AD), an irreversible neurodegenerative disorder that has caused the majority cases of dementia, wherein patients suffer progressive memory loss and cognitive function decline. Despite having no drugs for curing, early detection of AD allows the provision of preventive treatment to control the disease progression. The objective of this project is to develop a computer-aided system based on deep learning model to identify AD from cognitively normal and its early stage, mild cognitive impairment (MCI), using only structural MRI (sMRI). In this project, multiclass classification was applied. The dataset consisted of 3D T1-weighted brain sMRI images from the ADNI database containing 819 participants. A series of pre-processing methods were applied to the dataset; For example, skull stripping, bias field correction, pixel values normalisation, and data augmentation. HMRF tissue classifier was used to segment the brain MRI into 3 separate regions of grey matter, white matter, and cerebrospinal fluid. Axial brain images were extracted from the 3D MRI and being fed as input to the convolutional neural network (CNN) to perform multiclass classification of AD-CN-MCI. 3 different models were being experimented namely a CNN from scratch, VGG-16, and ResNet-50. The convolutional base of VGG-16 and ResNet-50 trained on ImageNet dataset were used as a feature extractor. Additionally, a new densely connected classifier was added on top of the convolutional base for performing classification. Using the 20% held out testing data, the performance of each model was reported and discussed. Among the 3 models, VGG-16 achieved the best testing performance with accuracy of 78.57%, precision of 73.94%, recall of 81.37%, and F1-score of 77.48%. Transfer learning technique allowed VGG-16 to achieve better performance despite a small number of data was being used. However, the best-performed VGG-16 has performed below average in comparison to previous works. Hence, limitations and possible solutions were outlined for future improvement.