GCAN: Generative Counterfactual Attention-guided Network for Explainable Cognitive Decline Diagnostics based on fMRI Functional Connectivity

TL;DR

GCAN tackles explainability in FC-based cognitive decline diagnostics by introducing counterfactual reasoning to identify brain regions predictive of SCD and MCI. It combines a Generator-Discriminator framework with an Atlas-aware Bidirectional Transformer (AABT) to reconstruct target FC and generate counterfactual attention maps that guide prediction. The model yields attention maps that align with known FC abnormalities and achieves superior diagnostic performance across HC/SCD/MCI tasks on hospital and ADNI datasets. This approach advances interpretable neuroimaging biomarkers for early Alzheimer's disease detection and provides a reproducible codebase.

Abstract

Diagnosis of mild cognitive impairment (MCI) and subjective cognitive decline (SCD) from fMRI functional connectivity (FC) has gained popularity, but most FC-based diagnostic models are black boxes lacking casual reasoning so they contribute little to the knowledge about FC-based neural biomarkers of cognitive decline.To enhance the explainability of diagnostic models, we propose a generative counterfactual attention-guided network (GCAN), which introduces counterfactual reasoning to recognize cognitive decline-related brain regions and then uses these regions as attention maps to boost the prediction performance of diagnostic models. Furthermore, to tackle the difficulty in the generation of highly-structured and brain-atlas-constrained FC, which is essential in counterfactual reasoning, an Atlas-Aware Bidirectional Transformer (AABT) method is developed. AABT employs a bidirectional strategy to encode and decode the tokens from each network of brain atlas, thereby enhancing the generation of high-quality target label FC. In the experiments of hospital-collected and ADNI datasets, the generated attention maps closely resemble FC abnormalities in the literature on SCD and MCI. The diagnostic performance is also superior to baseline models. The code is available at https://github.com/SXR3015/GCAN

Figures (4)

• Figure 1: The proposed Generative Counterfactual Attention-guided Network (GCAN) interpretatively enhances the diagnosis of cognitive decline by identifying key brain regions (referred to as 'counterfactual attention') involved in the transition between healthy, SCD, and MCI brain states, based on their functional connectivity (FC) matrix. This approach offers a bidirectional view of counterfactual attention that, once integrated with the original FC matrix, guides the classifier's focus to these regions throughout the learning process. The positive (${+}$) and negative (${-}$) signs symbolize two perspectives within the attention map, reflecting the inversion of causal relationship positions-source and target-in counterfactual inference.
• Figure 2: The proposed GCAN is illustrated as follows: The generator starts with a noisy source label input, $C^{s}_{n}$, using it to reconstruct the source label's FC. It also extracts disease-related causal information from the dataset's mean target label FC, $C^{t}_{r}$. This causal information is then leveraged to transform the reconstructed source label FC into the target label FC, $C^{t}_{g}$. On the other side, the discriminator employs components sensitive to both imaging features and neurodegenerative indicators to verify that $C^{t}_{g}$ not only mimics the FC characteristics accurately but also encapsulates distinct cognitive information differentiating it from $C^{s}_{r}$.
• Figure 3: The AABT mechanism involves the dynamic configuration of patch embedding and inverse patch embedding, contingent on the network's region count, facilitating both forward and backward processes. Here, $p_w$ and $p_h$ specify patch width and height, respectively, while $n_*$ indicates the network's region total.
• Figure 4: The generated target label FC and counterfactual attention in HC vs. MCI, HC vs. SCD, and SCD vs. MCI diagnostic task.