(P 175) INTERPRETABLE DEEP LEARNING BASED CLASSIFICATION OF SARCOMA IMMUNE SUBTYPES FROM H&E SLIDES FOR IMMUNOTHERAPY STRATIFICATION

Objective: A subset of sarcomas is immunogenic and responsive to immunotherapy. Identification of sarcoma immune subtypes (SICs) has previously required transcriptomic profiling. We aimed to develop an interpretable, AI‐driven pathology model to classify SICs directly from H&E‐stained whole‐slide images (WSIs), offering immune stratification based on histopathological features.

Methods: Using WSIs from The Cancer Genome Atlas (TCGA) with SICs annotated per Petitprez et al. (Nature 2020), we evaluated three embeddings and 5 aggregators in distinguishing SIC E (immunologically active) from non-E. Integrated Gradients (IG) analyses were performed to identify features contributing to SIC E classification.

Results: CONCHv1.5 with slide encoder achieved AUROC of 0.907 in distinguishing SIC E from non-E. Regions driving SIC E predictions had features such as high immune cell interconnectivity and tumor-immune proximity.

Conclusion: We developed a deep learning model capable of classifying SICs from H&E WSIs, offering scalable, sequencing-free identification of patients likely to benefit from immunotherapy.

Performance of models in distinguishing SIC E from non-E.