SCULPT: Semantic-aware causal prompt tuning for out-of-distribution detection of whole slide images

Out-of-distribution (OOD) detection in whole slide image (WSI) analysis presents significant challenges owing to the semantic ambiguity of cancer progression patterns. Traditional WSI analysis methods, which operate under closed-world assumptions, frequently fail to identify OOD samples effectively, potentially leading to critical errors in clinical settings. In this study, we propose Semantic-aware CaUsaL Prompt Tuning (SCULPT), a novel framework that leverages causal inference through front-door adjustment to address the issues of WSI OOD (WOOD) detection. SCULPT uses learnable causal prompts as mediator variables to establish explicit causal pathways between WSI inputs and classification outputs, mitigating the effects of tissue ambiguity. The framework comprises two key components: Semantic-aware Prompt Contrastive Learning (SPCL) for robust tissue prototype discovery through multi-scale semantic alignment, and Adaptive Causal Prompt Tuning (ACPT) for consistent and representative prototype learning through parallel hyper-factual and counterfactual prompt tuning. Experimental results on real-world datasets demonstrate that SCULPT achieves state-of-the-art performance in WOOD detection while also providing interpretable predictions through patch-level OOD maps and slide-level confidence scores. Moreover, SCULPT improves both performance and interpretability in-distribution classification tasks, representing a significant advancement in reliable and safe WSI analysis for clinical applications.