ESTRO 2025 - Abstract Book

S1053

Clinical – Head & neck

ESTRO 2025

Material/Methods: A dataset of 450 H&N cancer patients from TCGA-HNSC 1 was used, with each case including WSIs and ClinVars, split into 80% for cross-validation and 20% for testing. ClinVars included demographics, lifestyle, TNM, HPV and lymph nodes status. WSIs were processed using the Hierarchical Image Pyramid Transformer 2 (HIPT), a DL model to extract embeddings. ClinVars were encoded via TabTransformer 3 , another DL-based approach. These clinical embeddings were used to train a DeepSurv 4 model, also DL-based, to calculate survival risk scores for each patient. Patients were stratified into high- or low-risk groups, and this binary classification guided a CL 5 module to refine WSI embeddings further. The refined embeddings were then used to train a second DeepSurv model (Figure 1). Performance was assessed using the concordance index (C-index) on best variable subsets, validated via cross validation and test evaluations. For comparison, a baseline random survival forest (RSF) pipeline was implemented using only ClinVars.

Results: DeepSurv trained on baseline WSI embeddings from HIPT alone achieved a cross-validation C-index of 0.51 ± 0.03 and a test C-index of 0.49 ± 0.06. RSF using only ClinVars had cross-validation C-indexes up to 0.63 ± 0.03 and test C-indexes up to 0.79 ± 0.04. DeepSurv trained on ClinVar embeddings from TabTransformer without contrastive learning showed cross-validation C-indexes up to 0.66 ± 0.06 and test C-indexes up to 0.70 ± 0.05. Our DL approach with contrastive learning improved the cross-validation C-index to 0.77 ± 0.03 (Figure 2), indicating effective patient stratification from WSI refined embeddings. Although test set performance was lower (C-index of 0.50 ± 0.06), the significant cross-validation improvement and stratification with p-value below 0.001 suggest strong potential. The higher test C-index values observed for models trained with unimodal ClinVar may be attributed to the small test set size and inherent variability.