Learning with Domain-Knowledge for Generalizable Prediction of Alzheimer's Disease from Multi-site Structural MRI

Construct a generalizable model for the diagnosis of Alzheimer's disease (AD) is an important task in medical imaging. While deep neural networks have recently advanced classification performance for various diseases using structural magnetic resonance imaging (sMRI), existing methods often provide suboptimal and untrustworthy results because they do not incorporate domain-knowledge and global context information. Additionally, most state-of-the-art deep learning methods rely on multi-stage preprocessing pipelines, which are inefficient and prone to errors. In this paper, we propose a novel domain-knowledge-constrained neural network for automatic diagnosis of AD using multi-center sMRI. Specifically, we incorporate domain-knowledge into a ResNet-like architecture. We explicitly enforce the network to learn domain invariant and domain specific features by jointly training multiple weighted classifiers, so that pixel-wise predictive performance generalizes to unseen images. In addition, the network directly takes segmentation-free and patch-free images in original resolution as input, which offers accurate inference with global context information and accurate individualized abnormalities to further refines reproducible predictions. The framework was evaluated on a set of sMRI collected from 7 independent centers. The proposed approach identifies important discriminative brain abnormalities associated with AD. Experimental results demonstrate superior performance of our method compared to state-of-the-art methods.