Interpretable Gastric Cancer Survival Prediction via Concept Bottleneck Model

Gastric cancer is a leading cause of cancer-related mortality worldwide, making accurate prognosis essential for treatment decisions. Despite the promise of deep learning in whole slide image (WSI) analysis, its lack of interpretability limits clinical adoption. Traditional Concept Bottleneck Models (CBMs) provide transparency but rely heavily on labor-intensive manual annotations, hindering scalability. To address this, we propose a novel Concept Bottleneck framework specifically designed for WSI analysis to enable interpretable survival prediction of gastric cancer. Our framework combines a large language model and a vision-language model to automatically generate patch-level concept scores, reducing the need for manual annotation. Additionally, we introduce a transformer-based WSI-level concept score generator that captures structural interactions among tissue regions for effective score aggregation. Experimental results indicate that our method achieves performance comparable to state-of-the-art black-box WSI analysis models while maintaining interpretability. This work offers a reliable and transparent AI solution for clinical pathology, enhancing trust and applicability in real-world medical settings.