Biomarker Investigation Using Multiple Brain Measures from MRI Through Explainable Artificial Intelligence in Alzheimer's Disease Classification

Abstract

As the leading cause of dementia worldwide, Alzheimer's Disease (AD) has prompted significant interest in developing Deep Learning (DL) approaches for its classification. However, it currently remains unclear whether these models rely on established biological indicators. This work compares a novel DL model using structural connectivity (namely, BC-GCN-SE adapted from functional connectivity tasks) with an established model using structural magnetic resonance imaging (MRI) scans (namely, ResNet18). Unlike most studies primarily focusing on performance, our work places explainability at the forefront. Specifically, we define a novel Explainable Artificial Intelligence (XAI) metric, based on gradient-weighted class activation mapping. Its aim is quantitatively measuring how effectively these models fare against established AD biomarkers in their decision-making. The XAI assessment was conducted across 132 brain parcels. Results were compared to AD-relevant regions to measure adherence to domain knowledge. Then, differences in explainability patterns between the two models were assessed to explore the insights offered by each piece of data (i.e., MRI vs. connectivity). Classification performance was satisfactory in terms of both the median true positive (ResNet18: 0.817, BC-GCN-SE: 0.703) and true negative rates (ResNet18: 0.816; BC-GCN-SE: 0.738). Statistical tests (p < 0.05) and ranking of the 15% most relevant parcels revealed the involvement of target areas: the medial temporal lobe for ResNet18 and the default mode network for BC-GCN-SE. Additionally, our findings suggest that different imaging modalities provide complementary information to DL models. This lays the foundation for bioengineering advancements in developing more comprehensive and trustworthy DL models, potentially enhancing their applicability as diagnostic support tools for neurodegenerative diseases.

Keywords: Alzheimer’s disease; explainable artificial intelligence; magnetic resonance imaging; neuroimaging biomarkers; structural connectivity.

Figure 1

Implemented workflow for connectivity data extraction, AD classification, and explainability assessment performed on BC-GCN-SE. The processing steps used to derive the structural connectivity data, displayed in the orange panel, are described. The architecture of the implemented model (blue panel) comprised three GPC layers, an EP layer, an NP layer, and two fully connected layers. The outputs derived from the three convolutional layers of the model were processed using Grad-CAM (green panel) and then averaged. From the final heatmap, the contributions highlighted by each connection of each node of the HOA + AAL atlas were averaged to extract the Grad-CAM-based RV measure of each parcel. A connectogram representing both connectivity edges of the heatmap and color-coded Grad-CAM-based RV measures (circle perimeter) is displayed. These plots were created using the SPIDER-NET tool (i.e., Software Package Ideal (v1.0) for Deriving Enhanced Representations of brain NETworks) [36].

Figure 2

Implemented workflow for the AD classification and explainability assessment performed on ResNet18. The architecture of the implemented model (blue panel) comprised five convolutional layers: a GAP layer, three FC dense layers with ReLU activation, and a sigmoid activation function for the binary classification. The outputs derived from the last four convolutional layers were processed using Grad-CAM (green panel) and then averaged. The final heatmap was then multiplied by the binary masks underlying the HOA + AAL atlas parcels (the operation was indicated using an asterisk).

Figure 3

Boxplots of the Grad-CAM-based RV measures relative to the target parcels (specified on the x-axis) of the ResNet18 model. Results for the HC and AD subjects are represented using green and purple, respectively. For visualization purposes, all data were normalized between 0 and 1 using the minimum and the maximum Grad-CAM-based RV measures obtained across every parcel for every subject correctly classified using ResNet18. Lighter shades indicate parcels with significant Grad-CAM-based RV differences between AD and HC groups (Mann–Whitney or independent samples t-test).

Figure 4

Boxplots of the Grad-CAM-based RV measures relative to the target parcels (specified on the x-axis) of the BC-GCN-SE model. Results for the HC and AD subjects are represented using green and purple, respectively. For visualization purposes, all data were normalized between 0 and 1 using the minimum and the maximum Grad-CAM-based RV measures obtained across every parcel for every subject correctly classified using BC-GCN-SE. Lighter shades indicate parcels with significant Grad-CAM-based RV differences between AD and HC groups (Mann–Whitney or independent samples t-test).

Figure 5

Mean Grad-CAM-based RV measure of all parcels for the ResNet18 (a) and BC-GCN-SE (b) models. The AD and HC classes are reported separately. For visualization purposes, all data were normalized between 0 and 1 using the minimum and the maximum Grad-CAM-based RV measures obtained across every parcel for every subject correctly classified using the corresponding model. The brain lobe division is indicated through colored rectangles having different size proportionally to the number of parcels contained within each one. The Grad-CAM-based RV measure of every parcel is labeled using colored circles according to the following criterion: red indicates the 15% of parcels characterized by the highest Grad-CAM-based RV measure, yellow indicates the 15% of parcels characterized by the lowest Grad-CAM-based RV measure, gray indicates the remaining parcels. Plots were created using the SPIDER-NET tool [36]. Legend: Fro = frontal lobe; Ins = insular cortex; Lim = limbic lobe; Tem = temporal lobe; Par = parietal lobe; Occ = occipital lobe; SbC = subcortical structures; Ceb = cerebellum; BSt = brainstem.