S2308
Physics - Machine learning and AI algorithms
ESTRO 2026
tumour reached 0.61, and inflammatory infiltrates ≈ 0.58. Notably, the combined tumour + necrosis model reached 0.64, indicating that restricting segmentation to "cancer only" may be suboptimal; necrotic burden provides complementary prognostic signal in bladder cancer. The agent consistently identified subregions that improved C-index and produced concise, auditable rationales, supporting pathologist-driven inspection and discussion. Additional TCGA cohorts will be evaluated to assess generalisability.
oncologist to interactively query a slide (e.g., "show necrosis vs tumour vs stroma") and receive both compartment maps and class-specific survival utility. WSI-IA serves as a user interface for clinical professionals to work alongside AI, integrating seamlessly into radiotherapy and pathology workflows to support more explainable and informed treatment decisions. Instead of explaining model decisions post- hoc, we first partition the WSI into familiar tissue types and then quantify which tissues, individually or in combination, improve survival prediction. This supports clinically meaningful dialogue, auditing, and quality assurance grounded in tissue-level evidence. Material/Methods: Tiles are classified with CONCH[1], a pre-trained vision-language WSI tile classifier, into expert-defined classes: invasive urothelial carcinoma (tumour), normal urothelium, lamina propria, muscularis propria, perivesical adipose tissue, necrosis, inflammatory infiltrates, and blood vessels. Each tile is embedded with UNI-2[2], a WSI feature-embedding model that produces compact morphological representations. Embeddings are grouped by tissue class and aggregated into a class-level representation using a graph transformer. Survival models are trained per class and for selected class combinations on the TCGA-BLCA, a bladder urothelial carcinoma cohort, and performance is reported using the concordance index (C-index) with bootstrap resampling. The WSI-IA front-end follows the WSI-Agents paradigm[3]: a lightweight language model fine-tuned on pathology question-answer pairs and slide metadata, enabling natural-language queries. The agent produces compartment overlays, ranks tissue classes by prognostic contribution, and summarizes subregions driving survival predictions, aligning computational reasoning with pathologist cognition.
Conclusion: WSI-IA reframes interpretability around tissue compartments and measurable prognostic gain, moving beyond heatmaps toward pathologist- centric, queryable explanations. By revealing which tissues and which combinations drive survival prediction, it provides an auditable path for integrating computational pathology into oncology workflows and future radiotherapy decision support. References: 1.Lu MY, Chen B, Williamson DF, Chen RJ, Liang I, Ding T, Jaume G, Odintsov I, Le LP, Gerber G, Parwani AV. A visual-language foundation model for computational pathology. Nature medicine. 2024 Mar;30(3):863- 74. 2.Chen RJ, Ding T, Lu MY, Williamson DF, Jaume G, Song AH, Chen B, Zhang A, Shao D, Shaban M, Williams M. Towards a general-purpose foundation model for computational pathology. Nature medicine. 2024 Mar;30(3):850-62. 3.Lyu X, Liang Y, Chen W, Ding M, Yang J, Huang G, Zhang D, He X, Shen L. Wsi-agents: A collaborative multi-agent system for multi-modal whole slide image analysis. arXiv preprint arXiv:2507.14680. 2025 Jul 19. Keywords: Pathology AI, Explainability, Radiotherapy
Results: Tissue compartments expected to carry limited prognostic signal (e.g., muscularis propria, perivesical fat, benign mucosa) achieved C-index ≈ 0.55. Invasive
Made with FlippingBook - Share PDF online