Title: Explainable Diagnosis of Cognitive Decline via Multimodal Connectomes Using Brain-Atlas-Guided Generative Counterfactual Attention

Abstract

Accurate and interpretable diagnosis is critical for early risk assessment and intervention, particularly given the strong link between mild cognitive impairment (MCI), subjective cognitive decline (SCD), and the early stages of the Alzheimer’s disease continuum. While deep learning models based on connectomes have enhanced classification accuracy, they frequently offer scant insight into the functional and structural connectivity alterations associated with the disease. To address this, we introduce the Atlas-knowledge-guided Generative Counterfactual Attention-guided Network (GCAN), a framework designed for explainable cognitive decline diagnosis utilizing multimodal brain connectomes.

GCAN approaches diagnosis as a source-to-target counterfactual generation task. In this process, target-label connectomes are synthesized from source-label inputs, and the disparities between them are leveraged to create counterfactual attention maps. To maintain the topological integrity of the connectome, the framework employs an Atlas-aware Bidirectional Transformer (AABT), which executes network-level token encoding and decoding while adhering to brain-atlas constraints.

The methodology is expanded beyond functional connectivity (FC) to encompass joint modeling of functional and structural connectivity (SC). This extension facilitates counterfactual analysis of both complementary functional reorganization and structural topology shifts. Evaluations conducted on hospital-collected data and the ADNI dataset demonstrate that GCAN delivers competitive results in distinguishing between HC vs. SCD, HC vs. MCI, and SCD vs. MCI. The interpretability and reliability of the framework are further validated through visualizations, circular connectome analyses, CAM-based comparisons, ablation studies, and confidence interval assessments. Additionally, modality-specific FC and SC pre-trained classifiers supply target-state priors for the counterfactual generation process, remaining distinct from the downstream diagnostic classifier to avoid data leakage.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC