MMNet: A Lightweight Diagnostic Approach for Alzheimer's Disease Utilizing Small-Scale Data

Alzheimer's disease (AD), a systemic neurological disorder, is highly prevalent among the elderly, making early diagnosis the most effective form of treatment currently available. However, the manual screening process in its early stages is fraught with challenges, primarily due to the constraints of inadequate computing resources and the vast number of individuals affected by the disease. In this study, we introduce a streamlined modeling technique capable of diagnosing Alzheimer's Disease (AD) using small-scale datasets. Initially, Magnetic Resonance Imaging (MRI) or Positron Emission Tomography (PET) scans of AD patients are segmented and merged to create pseudo-RGB images. Subsequently, a compact network is designed to classify these images, facilitating feature extraction and processing. Notably, the inclusion of the Multi-Channel Attention (MCA) module within the compact network substantially enhances its feature extraction capabilities. Our model has been evaluated on a triple classification task utilizing a limited set of samples from the ADNI public dataset, achieving accuracies of 98.27% and 97.12% for PET and MRI images, respectively. Remarkably, these results were obtained while the model occupied merely 344M of memory.