Abstract
Deep learning models have demonstrated high accuracy in diagnosing Autism Spectrum Disorder (ASD) from functional MRI (fMRI) data, yet their "black-box" nature hinders clinical utility. This paper introduces two novel methods, Centralized Shapley Value Explanation (C-SVE) and Hierarchical Shapley Value Explanation (H-SVE), which integrate cooperative game theory with graph-based pruning to identify brain biomarkers efficiently. By leveraging the underlying functional connectivity graph, these methods overcome the exponential complexity of traditional Shapley values. Experimental results on MNIST and fMRI datasets show that C-SVE and H-SVE produce biologically plausible biomarkers and significantly outperform standard single-region importance metrics in model sensitivity analysis.
Introduction
While deep learning excels at pattern recognition in medical imaging, clinicians require interpretable explanations to identify biomarkers for diagnosis and treatment monitoring. In ASD research, identifying specific brain regions that drive a model's prediction is critical. However, standard interpretation methods often fail to account for the complex, networked interactions between brain regions, leading to "myopic" results that overlook key biomarkers.
The Challenge of Feature Dependency in Brain Networks
Most interpretation techniques rely on "leave-one-out" testing, which evaluates the impact of removing a single feature. This approach is inadequate for brain data where regions are conditionally dependent. If two regions are highly correlated, removing one may not alter the prediction, causing both to be undervalued.
Shapley Value Explanation (SVE) from cooperative game theory addresses this by considering all possible feature combinations (coalitions). However, calculating exact Shapley values for $N$ features requires $2^N$ computations, which is computationally impossible for high-dimensional brain atlases (e.g., 116+ regions).
Methodology: Graph-Based Shapley Value Explanation
The authors propose simplifying SVE by assuming that features only interact if they are connected within a functional brain graph. This graph is constructed using Pearson correlations of fMRI time series.
Centralized SVE (C-SVE)
This method restricts the calculation of a feature's marginal contribution to its immediate 1-step neighbors in the connectivity graph. By focusing on local interactions, C-SVE drastically reduces the number of coalitions that must be evaluated.
Hierarchical SVE (H-SVE)
This approach employs community detection to partition the brain into non-overlapping clusters. SVE is calculated within these communities, assuming negligible interaction between distant clusters. This hierarchical structure allows for efficient computation even in large-scale networks.
For larger neighborhoods or communities, a sampling-based Monte Carlo approach is used to estimate the Shapley values, ensuring the methods remain model-agnostic and computationally tractable.
Experimental Evaluation
The methods were validated using the MNIST digit dataset and a task-fMRI dataset of 130 children (82 ASD, 48 controls) performing a biological motion perception task.
MNIST Sanity Check
On MNIST, C-SVE and H-SVE produced smoother, more human-interpretable importance maps than standard ROI testing. Quantitatively, the proposed methods were superior at identifying features critical to the model's decision.
When the top 50% of features identified by C-SVE were corrupted, the model's prediction probability dropped by 0.720, compared to a 0.335 drop for the baseline single-region method.
ASD Biomarker Discovery
The identified ASD biomarkers—including the Putamen, Thalamus, and Superior Temporal Gyrus—align with established neurological literature. These regions are known to be involved in social communication and language processing. Functional decoding via Neurosynth further confirmed that these biomarkers correlate with high-level social concepts (e.g., "mind," "referential") and inversely correlate with lower-level skills (e.g., "language," "recognition").
Discussion and Conclusion
By integrating graph theory with cooperative game theory, C-SVE and H-SVE provide a computationally efficient and biologically grounded framework for interpreting deep learning models. These methods successfully move beyond "myopic" feature testing, capturing the synergistic interactions of brain networks to discover meaningful ASD biomarkers. Future work will focus on disentangling specific pairwise interactions and testing these methods across diverse brain atlases.
Summary updated