Concept-Level Causal Explanation Method for Brain Function Network Classification

Jinduo Liu, Feipeng Wang, Junzhong Ji, Dongxiao He, Bingdao Feng, Zhen Wang

Abstract

Deep learning models have demonstrated significant potential in diagnosing brain diseases using functional Magnetic Resonance Imaging (fMRI). However, their "black-box" nature limits clinical utility. We propose the Concept-Level Causal Explanation Method (CLCEM), a novel framework that maps abstract neural network features to clinically relevant Regions of Interest (ROIs). By integrating a causal estimator and a Causal Concept Loss (CCL), CLCEM ensures that each extracted feature corresponds to a specific brain region. Our results on the ABIDE I dataset for Autism Spectrum Disorder (ASD) show that CLCEM achieves a state-of-the-art accuracy of 68.37% while providing transparent, clinically aligned reasoning paths.

Introduction

Deep learning has become a powerful tool for classifying Brain Functional Networks (BFNs)—matrices derived from fMRI data that represent the communication patterns between different brain regions. Despite their performance, these models often lack interpretability. Existing explanation methods, such as saliency maps or Grad-CAM, provide "pixel-level" explanations that highlight specific cells in a matrix, which are often clinically meaningless. Furthermore, standard concept-based explainers are typically designed for natural images and fail to translate to the abstract nature of BFNs. There is an urgent need for methods that translate neural network features into the language of medicine: Regions of Interest (ROIs).

Methodology

The CLCEM framework consists of three primary phases designed to ensure both performance and interpretability.

1. Concept Extraction

The model employs a causal estimator, constructed with non-linear Multilayer Perceptrons, to evaluate the causal relationship between deep features and specific ROIs. A key innovation is the Causal Concept Loss (CCL), which regularizes the model to bind each extracted feature to exactly one brain region, ensuring a clear mapping between AI features and anatomical locations.

ROI-based feature mapping in the CLCEM framework

2. Concept Aggregation and Decision

Since multiple features may relate to the same brain region, CLCEM uses "location connections" to group these features. This allows the model to calculate the total contribution of each ROI to the final diagnostic output. Instead of a final "black box" layer, the model produces a probability based on the summed contributions of these ROIs, effectively turning the classification process into a transparent tally of evidence.

The CLCEM architecture illustrating the flow from concept aggregation to the final diagnostic decision.

Experimental Results

The study utilized the ABIDE I (Autism Brain Imaging Data Exchange) database, comprising 1,035 individuals (505 ASD, 530 controls). Using the Automated Anatomical Labeling (AAL) template to define 90 ROIs, CLCEM achieved a classification accuracy of 68.37%, outperforming baseline explainable methods such as LIME (66.09%) and Grad-CAM (65.80%).

Key Insights

  • Causality vs. Correlation: By using Structural Equation Modeling (SEM), CLCEM identifies ROIs that actively drive the model's decision rather than those that merely correlate with the output.
  • Feature Pruning: Interestingly, using only the top 30 ROIs identified by CLCEM resulted in better performance than using all 90 ROIs, suggesting that irrelevant brain regions introduce noise that hinders deep learning models.
  • Clinical Alignment: The ROIs identified as most significant—including the Right Lingual Gyrus and Right Superior Temporal Gyrus—align with established medical literature regarding ASD symptoms, such as impaired social cognition and visual processing.

Conclusion

CLCEM provides a robust framework for making deep learning models in neuroscience transparent. By mapping abstract features to ROIs and utilizing causal learning, it offers a "reasoning path" that aligns with clinical knowledge. Future work will extend this binary classification model to multi-class diagnosis and incorporate non-visual clinical data to further enhance interpretability.

Summary updated

BibTeX
@inproceedings{liu2024concept,
  author = {Jinduo Liu and Feipeng Wang and Junzhong Ji and Dongxiao He and Bingdao Feng and Zhen Wang},
  title = {{Concept-Level Causal Explanation Method for Brain Function Network Classification}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=MzFQiP45Zf},
}
Figure