Abstract
Using deep models to classify brain functional networks (BFNs) for the auxiliary diagnosis and treatment of brain diseases has become increasingly popular. However, the unexplainability of deep models has seriously hindered their applications in computer-aided diagnosis. In addition, current explanation methods mostly focus on natural images, which cannot be directly used to explain the deep model for BFN classification. In this paper, we propose a concept-level causal explanation method for BFN classification called CLCEM. First, CLCEM employs the causal learning method to extract concepts that are meaningful to humans from BFNs. Second, it aggregates the same concepts to obtain the contribution of each concept to the model output. Finally, CLCEM adds the contribution of each concept to make a diagnosis. The experimental results show that our CLCEM can not only accurately identify brain regions related to specific brain diseases but also make decisions based on the concepts of these brain regions, which enables humans to understand the decision-making process without performance degradation.
Introduction
In the field of medical AI, "black box" models present a significant barrier to clinical adoption. While deep learning excels at identifying patterns in brain activity, clinicians require transparent reasoning to trust automated diagnoses. This is particularly critical for Brain Functional Networks (BFNs), which map complex communications between brain regions. Traditional explanation methods, such as saliency maps, often fail because they highlight abstract "pixels" in connectivity matrices that lack biological significance to a clinician.
Methodology
The proposed Concept-Level Causal Explanation Method (CLCEM) transforms deep learning models into "glass boxes" by grounding decisions in human-understandable medical concepts—specifically, Regions of Interest (ROIs). The framework operates in three distinct phases:
1. Concept Extraction
CLCEM utilizes a Causal Estimator, constructed from non-linear Multilayer Perceptrons, to determine the causal relationship between deep features and specific brain regions. A novel Causal Concept Loss (CCL) is implemented to force the model to extract features with strong causal links to individual ROIs, effectively binding mathematical features to biological concepts.
2. Concept Aggregation and Decision
As shown in the framework overview, the model groups features by their corresponding ROIs using "location connections." This aggregates individual contributions into a unified score for each concept. The final diagnostic decision (e.g., "Patient" vs. "Healthy") is a simple sum of these ROI contributions, ensuring the logic is entirely transparent and additive.
Experimental Results
The researchers evaluated CLCEM using the ABIDE I (Autism Brain Imaging Data Exchange) database, comprising 1,035 individuals. fMRI data was mapped to the AAL (Automated Anatomical Labeling) template, defining 90 specific ROIs.
CLCEM achieved a classification accuracy of 68.37%, outperforming both the baseline (65.02%) and other state-of-the-art explanation methods like LIME and Grad-CAM. It also maintained high performance across other metrics, including a Balanced Accuracy of 68.16% and an AUC of 71.75%. Crucially, the ROIs identified as most significant for Autism Spectrum Disorder—such as the Lingual Gyrus and Superior Temporal Gyrus—align with established medical literature, validating the model's clinical relevance.
Conclusion
CLCEM demonstrates that interpretability does not require a performance trade-off. By constraining models to focus on biologically relevant concepts, CLCEM reduces noise and improves generalization. Future work will expand this framework to multi-class diagnosis and incorporate multimodal data like structural MRI to further enhance clinical utility.
Summary updated