[Paper Review] Biomarker Investigation using Multiple Brain Measures from MRI through XAI in Alzheimer's Disease Classification

Davide Coluzzi, Valentina Bordin+5 more|arXiv (Cornell University)|May 3, 2023

Functional Brain Connectivity Studies4 citations

TL;DR

This study proposes two deep learning models—ResNet18 for structural MRI and BC-GCN-SE for brain connectivity matrices—combined with Explainable AI (XAI) via Grad-CAM to identify biomarkers in Alzheimer’s disease. It demonstrates that both models localize attention on clinically relevant brain regions, such as the medial temporal lobe and default mode network, with median true positive rates of 0.817 and 0.703, respectively, enhancing model interpretability and clinical trust.

ABSTRACT

Alzheimer's Disease (AD) is the world leading cause of dementia, a progressively impairing condition leading to high hospitalization rates and mortality. To optimize the diagnostic process, numerous efforts have been directed towards the development of deep learning approaches (DL) for the automatic AD classification. However, their typical black box outline has led to low trust and scarce usage within clinical frameworks. In this work, we propose two state-of-the art DL models, trained respectively on structural MRI (ResNet18) and brain connectivity matrixes (BC-GCN-SE) derived from diffusion data. The models were initially evaluated in terms of classification accuracy. Then, results were analyzed using an Explainable Artificial Intelligence (XAI) approach (Grad-CAM) to measure the level of interpretability of both models. The XAI assessment was conducted across 132 brain parcels, extracted from a combination of the Harvard-Oxford and AAL brain atlases, and compared to well-known pathological regions to measure adherence to domain knowledge. Results highlighted acceptable classification performance as compared to the existing literature (ResNet18: TPRmedian = 0.817, TNRmedian = 0.816; BC-GCN-SE: TPRmedian = 0.703, TNRmedian = 0.738). As evaluated through a statistical test (p < 0.05) and ranking of the most relevant parcels (first 15%), Grad-CAM revealed the involvement of target brain areas for both the ResNet18 and BC-GCN-SE models: the medial temporal lobe and the default mode network. The obtained interpretabilities were not without limitations. Nevertheless, results suggested that combining different imaging modalities may result in increased classification performance and model reliability. This could potentially boost the confidence laid in DL models and favor their wide applicability as aid diagnostic tools.

Motivation & Objective

• To improve diagnostic accuracy in Alzheimer’s disease using deep learning on multimodal MRI data.
• To enhance clinical trust in AI models by applying Explainable AI (XAI) techniques to interpret model decisions.
• To validate model predictions against known pathological brain regions using Grad-CAM interpretability analysis.
• To assess whether combining structural and functional brain measures improves classification performance and reliability.
• To identify the most relevant brain parcels for Alzheimer’s classification through statistical ranking of Grad-CAM saliency maps.

Proposed method

• Trained a ResNet18 model on T1-weighted structural MRI scans to classify Alzheimer’s disease.
• Developed a BC-GCN-SE model to process brain connectivity matrices derived from diffusion MRI data.
• Applied Grad-CAM to generate class activation maps for interpretability across 132 brain parcels from the Harvard-Oxford and AAL atlases.
• Mapped saliency scores from Grad-CAM to anatomical regions and compared them to known pathological zones in Alzheimer’s disease.
• Used statistical testing (p < 0.05) and ranking of top 15% most relevant parcels to evaluate model adherence to domain knowledge.
• Evaluated model performance using median true positive rate (TPR) and true negative rate (TNR) across cross-validation folds.

Experimental results

Research questions

• RQ1 Which brain regions do deep learning models attend to when classifying Alzheimer’s disease from MRI data?
• RQ2 To what extent do the attention maps of these models align with known pathological regions in Alzheimer’s disease?
• RQ3 How does combining structural MRI and brain connectivity data affect classification performance and interpretability?
• RQ4 Can XAI techniques like Grad-CAM effectively validate and enhance trust in deep learning models for neurodegenerative disease classification?
• RQ5 Do the most salient brain parcels identified by Grad-CAM correspond to clinically relevant networks such as the default mode network or medial temporal lobe?

Key findings

• The ResNet18 model achieved a median true positive rate (TPR) of 0.817 and median true negative rate (TNR) of 0.816, indicating strong classification performance.
• The BC-GCN-SE model achieved a median TPR of 0.703 and median TNR of 0.738, demonstrating moderate but clinically relevant accuracy.
• Grad-CAM analysis revealed that both models focused attention on the medial temporal lobe and the default mode network—key regions implicated in Alzheimer’s pathology.
• Statistical testing confirmed that the top 15% of most relevant brain parcels significantly overlapped with known pathological regions (p < 0.05).
• The interpretability analysis demonstrated that model predictions were not arbitrary but aligned with established neuroanatomical knowledge of Alzheimer’s disease.
• The results suggest that multimodal MRI integration with XAI can enhance model reliability and support clinical adoption of deep learning in neurodegenerative disease diagnosis.