A Systematic Evaluation of Saliency Methods for 3D MRI-Based Alzheimer’s Disease Classification

Shubham Joshi, Millie Pant, Kusum Deep

Keywords: Alzheimer, Neuroimaging, Class Activation Maps, Interpretable AI, Explainable AI

Abstract:

Structural MRI is essential to the assessment of Alzheimer’s disease (AD), yet the limited
interpretability of deep neural networks continues to hinder their broader clinical adoption.
Class Activation Map (CAM) approaches are used to visualize the decision making of deep
neural networks, but their behaviour in three-dimensional neuroimaging contexts is still not
well characterized. This study provides a systematic assessment of six saliency methods
viz. Grad-CAM, Grad-CAM++, EigenCAM, LayerCAM, ScoreCAM, and ReciproCAM
with a backbone 3D ResNet-18 classifier trained on 1540 T1-weighted MRI scans from the
Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. We also present an anatom
ically based evaluation framework that quantifies regional activation distributions through
predefined volumetric regions of interest (ROIs) encompassing five structures associated
with Alzheimer’s disease (AD). CAM energy fractions are calculated by region for Cog
nitively Normal (CN), Mild Cognitive Impairment (MCI), and Alzheimer’s Disease (AD)
cohorts. In addition to these spatial analyses, we conduct a comprehensive quantitative
evaluation employing deletion, insertion, and ROAD faithfulness metrics on the complete
test set. Our results show significant variability between CAM approaches in faithfulness
and anatomical localisation, suggesting that qualitative heatmaps by themselves might not
give a complete picture of network reasoning. The study offers a reproducible benchmark of
six CAM methods for evaluating explainability approaches in 3D neuroimaging, and high
lights considerations for selecting reliable saliency tools for MRI-based AD diagnosis. Our
code is available at https://github.com/sjiitr/Benchmarksaliency_alzheimer.git.

Primary Subject Area: Interpretability and Explainable AI

Secondary Subject Area: Application: Neuroimaging

Registration Requirement: Yes

Reproducibility: https://github.com/sjiitr/Benchmarksaliency_alzheimer.git

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 16