CDR-CNN: Convolutional Neural Networks for Alzheimer’s Disease Severity Prediction with Clinical Dementia Rating

Precisely identifying the severity of Alzheimer’s using MRI is crucial for enhancing diagnostic accuracy and improving patient care. This study presents a novel convolutional neural network designed to classify four classes of Alzheimer: none, very mild, mild, and moderate. The dataset, comprising 80,000 MRI scans from 461 participants, tackles the issue of substantial class skew, with a predominance of scans in the ‘none’ category. The proposed method deviates from conventional approaches by preserving ordinal information in predictions, as opposed to employing one-hot encoding for categorical classification. This study examines the effectiveness of ordinal classification relative to categorical classification, explores strategies for addressing class imbalance, and assesses the performance of various pre-trained model bases. This research aims to enhance diagnostic accuracy for AD, providing valuable insights for the medical, neuroscientific, and data science communities working with MRI and other medical imaging data.