Deep Learning for Alzheimer ’ s Disease Diagnosis : Global Prediction and Weakly-Supervised Localization

Synopsis Most diffusion MRI approaches rely on comparably long scan time and a suboptimal processing pipeline with handcrafted physical/mathematical representations. They can be outperformed by recent handcrafted-representation-free methods. For instance, qspace deep learning (q-DL) allows unprecedentedly short scan times and optimized voxel-wise tissue characterization. We reformulate qDL such that it estimates global (i.e. scan-wise rather than voxel-wise) information. We use this formulation to distinguish Alzheimer’s disease (AD) patients from healthy controls based solely on raw q-space data without handcrafted representations such as DTI. Classification quality is very promising. Weakly-supervised localization techniques indicate that the neural network attends to ADrelevant brain areas.