Explainable AI for Alzheimer’s Prediction Diagnosis: Developing an Interpretable Deep Learning Pipeline with 3D Visualization

Sammanfattning

Deep learning methods are growing in complexity and inherently trading interpretability for accuracy. While accuracy is important, equal emphasis should be placed on explaining the model’s internal workings to justify its decisions. In this work, an explainable pipeline was developed for a CNN model for predicting Alzheimer’s diagnosis. Specifically, it determines the brain region’s relevance in classifying cognitively normal (CN) and Alzheimer’s disease (AD) using 3D structural MRI images. The XAI methods used in the implementation were DeepLIFT, Layer Gradient x Activation, Occlusion, and XGrad-CAM to obtain 3D visual explanations of discriminative regions derived from the model’s internal workings. The regional relevance was assessed in a cross-sectional comparison between CN and AD subjects and a longitudinal comparison of subjects transitioning from CN to AD, tracking the relevance in time. A correlation analysis between regional relevance and mean standardized uptake value ratio (SUVR) was also conducted to investigate structural and bioactive overlap. The results showed that the model uses a relatively concentrated region centered around the temporal lobe to predict CN subjects and a much larger and disipated area for the AD predictions. The longitudinal comparison found the model attributing increasing importance to the caudate, putamen, and pallidum. The Occlusion relevance of the hippocampus and ventricles showed a positive correlation with the Amyloid PET mean SUVR. This supports the connection between anatomical brain changes and neuropathologic biomarkers used to diagnose Alzheimer’s disease.