CausalX-Net: a causality-guided explainable segmentation network for brain tumors

초록

Brain tumors represent a significant health challenge in India, with approximately 28,000 new cases diagnosed annually. Conventional deep learning approaches for MRI-based segmentation often struggle with irregular tumor boundaries, heterogeneous intensity patterns, and complex spatial relationships, resulting in limited clinical interpretability despite high numerical accuracy. This study introduces CausalX-Net, a causality-guided explainable segmentation network for brain tumor analysis from multi-modal MRI. Unlike purely correlation-based models, CausalX-Net leverages structural causal modeling and interventional reasoning to identify and quantify the causal influence of imaging features and spatial regions on segmentation outcomes. Through counterfactual analysis, the framework can provide clinically relevant "what-if" explanations, such as predicting changes in tumor classification if specific modalities, regions, or features are altered. Evaluated on the BraTS 2021 dataset, CausalX-Net achieved a Dice Similarity Coefficient of 92.5%, outperforming state-of-the-art CNN-based baselines by 4.3% while maintaining competitive inference efficiency. Furthermore, causal attribution maps and intervention-based sensitivity analyses enhance trust and transparency, offering radiologists actionable insights for diagnosis and treatment planning. This research demonstrates that integrating causal inference into segmentation not only improves accuracy but also delivers interpretable, decision-supportive explanations, representing a significant step toward transparent and reliable AI-assisted neuroimaging in clinical settings.

키워드