Causal Pathways in Multi-Modal Imaging AI: From Attribution Maps to Interventional Explanations

While Jin et al. (2022, Source: AAAI) highlight that current heatmap-based explanations often fail to meet clinical needs in multi-modal imaging, and Baron (2023) discusses the limitations of counterfactuals in providing genuine causal understanding, this idea proposes synthesizing these insights. Instead of only displaying which regions of which modalities are “important” (attribution), the research would develop and validate methods that simulate causal interventions — e.g., “What if the edema region in FLAIR were less intense?” — and show their effect on predictions (such as tumor grade). This causal, intervention-based XAI could help clinicians understand not just where, but how and why the model’s decision would change, aligning explanations with clinical reasoning and potentially uncovering spurious correlations or biases. This approach reframes the conceptual framework of XAI in biomedical imaging, borrowing from the causal inference literature (Shmueli et al., 2025; Baron, 2023), and could fundamentally improve model trustworthiness and utility in clinical workflows.

References:

1. Evaluating Explainable AI on a Multi-Modal Medical Imaging Task: Can Existing Algorithms Fulfill Clinical Requirements?. Weina Jin, Xiaoxiao Li, Ghassan Hamarneh (2022). AAAI Conference on Artificial Intelligence.
2. From What Ifs to Insights: Counterfactuals in Causal Inference vs. Explainable AI. Galit Shmueli, David Martens, Jaewon Yoo, Travis Greene (2025). arXiv.org.
3. Explainable AI and Causal Understanding: Counterfactual Approaches Considered. S. Baron (2023). Minds and Machines.