Comparison of Explainable AI Models for MRI-based Alzheimer's Disease Classification

Abstract

Deep learning models based on convolutional neural networks (CNNs) have been used to classify Alzheimer's disease or infer dementia severity from 3D T1-weighted brain MRI scans. Here, we examine the value of adding occlusion sensitivity analysis (OSA) and gradient-weighted class activation mapping (Grad-CAM) to these models to make the results more interpretable. Much research in this area focuses on specific datasets such as the Alzheimer's Disease Neuroimaging Initiative (ADNI) or National Alzheimer's Coordinating Center (NACC), which assess people of North American, predominantly European ancestry, so we examine how well models trained on these data generalize to a new population dataset from India (NIMHANS cohort). We also evaluate the benefit of using a combined dataset to train the CNN models. Our experiments show feature localization consistent with knowledge of AD from other methods. OSA and Grad-CAM resolve features at different scales to help interpret diagnostic inferences made by CNNs.

Keywords: Alzheimer’s Disease; Deep Learning; Grad-CAM; Magnetic Resonance Imaging; Occlusion Sensitivity Analysis.

Figure 1.

DenseNet Architecture with OSA Model.

Figure 2.

Comparison of explainable models for controls (CN) and participants with dementia in the NACC test dataset. The first column shows the original T1w scan’s middle axial slice (40), for a random subject in the test set. The top row shows a subject with dementia, and the bottom row shows a healthy control participant. The middle column shows the OSA group averages, and the last column shows the corresponding Grad-CAM group averages for 100 participants of the NACC test dataset. Grad-CAM maps detect features at a finer spatial scale than OSA; the OSA maps appear ‘blocky’ as the occlusion kernels do not overlap. If overlap were allowed, these maps would still be spatially smooth. By contrast, Grad-CAM maps show in deep blue the ventricles, which show strong group differences between patients and controls.

Figure 3.. Atrophy map comparing an AD patient group to healthy controls, using tensor-based morphometry in the ADNI dataset.

Although maps of salient features are hard to validate without ground truth on the location of the disease, complementary image analysis methods, such as TBM, show ventricular expansion and subcortical atrophy. These key signs of dementia are used by radiologists for diagnosis. (Reproduced from Hua et al. [20]).