An End-to-End 3D ConvLSTM-based Framework for Early Diagnosis of Alzheimer's Disease from Full-Resolution Whole-Brain sMRI Scans

Alzheimer's Disease (AD) is the most prevailing form of dementia, killing more people than prostate and breast cancers combined. Structural Magnetic Resonance Imaging (sMRI) is widely used for the analysis of progressive brain aggravation and its clinical utility in discriminating AD is well established. Even if an effective cure does not exist yet, early detection is fundamental for slowing down the worsening of symptoms. Thus, the aim of the present work is to propose an end-to-end 3D Convolutional Long Short-Term Memory (ConvLSTM)-based framework for early diagnosis of AD from full-resolution whole-brain sMRI scans. The proposed framework was applied to 427 full-resolution whole-brain sMRI scans belonging to both OASIS and ADNI databases in order to provide a less dataset-specific approach. Results show that our framework is performing well in discriminating AD from Cognitively Normal (CN) patients, reaching a classification accuracy of 86%, sensitivity of 96%, f1-score of 88% and AUC of 93% on the test data. The tests were performed on a scalable GPU cloud service and are publicly available to guarantee reproducibility. Since the proposed framework performs well without domain-specific knowledge from AD as well as computationally-costly processes such as segmentation, it can be applied to other mental disorders using whole-brain sMRI scans as input data.