CT-Integrated Pathophysiology-Guided, Clinically Causal Cognition-Led Interpretable Lesion Context Characterization and Risk Reasoning in Lung MWA

Meng Li¹Yongzhao Li²Xiangming Wang¹Hui Feng¹Yang Li¹Xinbo Liu^3*

• ¹Department of Radiology,The Fourth Hospital of Hebei Medical University, Shijiazhuang, China
• ²Center of Treatment of Myasthenia Gravis, People’s Hospital of Shijiazhuang affiliated to Hebei Medical University, Shijiazhuang, China
• ³Department of Thoracic Surgery,The Fourth Hospital of Hebei Medical University, Shijiazhuang, China

Background: Accurate preoperative prediction of complication risks in microwave ablation (MWA) for non-small cell lung cancer (NSCLC) is critical yet challenging due to the complex anatomical context. Beyond feature heatmaps, clinicians need decision rationales that align with pathophysiology. We therefore frame the task as an interpretable lesion-context characterization driven by clinical causal cognition, where imaging-derived concepts (e.g., tumor–pleura distance, vascular proximity, pleural adhesions) are organized into traceable causal pathways supporting downstream risk reasoning. . Objective: This study aimed to validate that a previously developed multi‑ task, attention‑ enhanced framework can be re-expressed through a concept bottleneck and lesion-context graph to provide pathophysiology-consistent, interpretable reasoning for lesion characterization and to support downstream prediction of pneumothorax, hemorrhage, and pleural reactions. Methods: This study retrospectively analyzed 184 NSCLC‑ MWA cases with paired thin‑ slice CT and expert annotations. Our architecture integrates a multi‑ scale cross‑ spatial attention‑ enhanced 3D U‑ Net and multi‑ task heads. On top of the existing pipeline, we introduce a concept bottleneck layer (tumor volume/shape, minimum tumor–pleura distance, vascular proximity, pleural adhesions), a post‑ hoc lesion‑ context graph with message passing for explanation, and a causal‑ consistency evaluation suite (directionality/monotonicity/sign tests, cross‑ subgroup invariance, counterfactual sensitivity). Human–AI collaboration was assessed by radiologists' usability scores for the generated causal pathways. No retraining was required. Results: The framework maintained anatomical segmentation performance (Dice: tumor 0.878, vessels 0.851, adhesions 0.863) and multi-task AUCs (0.903/0.871/0.847 for pneumothorax/hemorrhage/pleural reactions). Causal-consistency showed expected negative monotonicity between tumor-pleura distance and pneumothorax risk (Kendall's τ=+−0.61; violation rate 9%), positive between vascular proximity and hemorrhage (τ=+0.57; violations 12%), and positive between adhesions and pleural reactions (τ=+0.49; violations 16%). Counterfactual geometric edits (±5 mm pleural distance; ±3 mm vascular proximity) produced direction-consistent risk changes in 86% and 83% of cases, respectively. Conclusion: By translating imaging features into clinically meaningful, traceable causal frameworks, our hybrid system aligns AI logic with clinical cognition for interpretable lesion-context diagnosis and reliable downstream risk reasoning. This "from black-box to clarity" shift improves trustworthiness and potential clinical utility.