Anatomy-Aware Gating Network for Explainable Alzheimer's Disease Diagnosis

Hongchao Jiang, Chunyan Miao

Abstract: Structural Magnetic Resonance Imaging (sMRI) is a non-invasive technique to get a snapshot of the brain for diagnosing Alzheimer’s disease. Existing works have used 3D brain images to train deep learning models for automated diagnosis, but these models are prone to exploit shortcut patterns that might not have clinical relevance. We propose an Anatomy-Aware Gating Network (AAGN) which explicitly extracts features from various anatomical regions using an anatomy-aware squeeze-and-excite operation. By conditioning on the anatomy-aware features, AAGN dynamically selects the regions where atrophy is most discriminative. Once trained, we can interpret the regions selected by AAGN as explicit explanations for a given prediction. Our experiments show that AAGN selects regions well-aligned with medical literature and outperforms various convolutional and attention architectures. The code is available at https://github.com/hongcha0/aagn.