Analysis of CNN models in classifying Alzheimer's stages: comparison and explainability examination of the proposed separable convolution-based neural network and transfer learning models

Dementia is a condition that affects brain functions and usually occurs with age, resulting in a decrease in cognitive functions (such as thinking, remembering, and decision-making). The most common type of dementia is Alzheimer's disease. Early diagnosis of Alzheimer's disease is essential to ensure benefit from treatment options, help plan the future, and facilitate the management of symptoms. Significant advances have been made in AI-assisted MRI analysis for early diagnosis of Alzheimer's disease. This study proposed a separable convolution-based neural network called DEMxNET to classify four different Alzheimer's stages. The proposed model was compared with nine different state-of-the-art transfer learning models named VGG16, VGG19, InceptionV3, Xception, DenseNet201, MobileNetV2, ResNetRS101, NASNetMobile, and ConvNeXtBase. The Synthetic Minority Oversampling Technique (SMOTE) was used to solve the class imbalance problem in the dataset. The proposed DEMxNET model achieved the highest performance compared to other models and reached an accuracy rate of 99.79%. In addition to artificial intelligence-supported diagnostic methods, the explainability and understandability of how the diagnosis is made are also essential for patients and healthcare professionals. In this respect, the explainability and the performance of the DEMxNET model were considered. The explainability of the model was ensured using the LIME method.