A novel explainable neural network for Alzheimer’s disease diagnosis

Highlights：

• We propose a novel 3D interpretable model, dubbed MAXNet, which can effectively aggregate multi-scale features for Alzheimer’s disease detection and learn latent features that are representative to each volume’s label.

• We present a novel high-resolution visualization method, termed Highresolution Activation Mapping (HAM), that produces high-resolution visual explanations for the precise localization of disease areas through aggregating the attentional representations from multi-resolution responses in parallel.

• We propose a Prediction-basis Creation and Retrieval (PCR) module, which leverages latent representations to collect similar reference samples as visual evidence for the case analysis of Alzheimer’s disease.