A Deep Transfer Learning Model for Classifying Alzheimer's Disease Using MRI Images

Alzheimer's disease (AD) affects the cells of the human brain and leads to the development of dementia. The progressive brain cells degeneration results in the inability to carry out everyday duties. The therapy for this condition is currently insufficiently developed. Nevertheless, detecting the AD at an early stage can help control its transmission. To detect AD at an early stage using brain Magnetic Resonance Imaging (MRI), it is necessary to create an automated system that can identify and categorize individuals with dementia. In this research, a deep learning (DL) model is developed for the classification of AD based on MRI pictures. The research model was created by applying a pre-trained transfer learning model known as ResNeXt-101. The OASIS-3 dataset of MRI images is initially gathered and utilized for both training the model and evaluating its performance. The dataset comprises 2842 MRI pictures pertaining to AD and its different phases. As part of the preprocessing stage, all the images are resized and cropped to a uniform size of 224×224 pixels. Furthermore, they are enhanced utilizing the contrast enhancement technique. The ResU-Net segmentation method is utilized following the preprocessing phase. The ResNeXt-101 model is employed for the purpose of classification the segmented images. The MRI pictures are divided into training and testing sets, with 75% allotted to train the model and 25% allocated to evaluate the model. The classification is conducted using a binary approach, categorizing the data as either the presence or absence of AD. The research model's performance was evaluated using metrics such as accuracy, specificity, sensitivity, f1-score, and precision. The research model achieved a training accuracy of 99.67%, sensitivity of 99.54%, precision of 99.72%, specificity of 99.41%, and a f1-score of 99.60%. The model achieved an accuracy of 99.38%, a sensitivity of 98.29%, a precision of 99.40%, a specificity of 99.24%, and a f1-score of 99.36% in the performance evaluation. The test results were compared to the existing models for validation. The research model has demonstrated superior performance when compared to all the other models.