Prospective-retrospective HeCOG phase III trial validation of Polaris TME for the prediction of TILs from whole-slide images in high-risk early breast cancer.

Abstract

557 Background: Tumor-infiltrating lymphocytes (TILs) are an established prognostic and predictive biomarker in early breast cancer. However, manual histopathologic scoring can introduce interobserver variability and limit throughput in analyzing at scale in routine practice. Recent advances in artificial intelligence (AI) offer the potential to standardize TILs quantification and enhance reproducibility. Our aim is to evaluate whether deep learning-based tumor microenvironment (TME) quantification from digitized hematoxylin- and eosin (H&E)-stained slides can accurately classify TILs and assess their prognostic role in patients with early breast cancer. Methods: We conducted a prospective-retrospective external validation of Polaris TME, a deep learning model for TME quantification. The analysis included patients with high-risk early breast cancer treated with adjuvant contemporary dose-dense chemotherapy, spanning all histological subtypes, enrolled in four HeCOG trials conducted between 1997 and 2008 (3 randomized and 1 observational). Manual TILs scoring had been centrally performed by pathologists according to International TILs Working Group guidelines, enabling direct comparison between manual and AI-derived scoring. The Polaris foundation model was used for feature extraction, while model performance was assessed using correlation for continuous TILs scores using the Spearman ρ and discrimination for clinically defined TILs categories (thresholds: 5% and 50% using the Area Under the Curve (AUC) with 95% CI. The prognostic value of the model was assessed using the log-rank test for invasive disease-free survival (iDFS). Results: Overall, 1,214 patients with whole slide images, TILs quantification and follow-up data were included; median age 53 [22-79]; 638 (63%) were hormone receptor-positive/HER2-negative, 394 (32%) HER2-positive, 176 (14%) triple-negative breast cancer (TNBC), and 6 patients of unknown subtype. Median follow-up was 116.59 (53.45 –157.06) months. When blindly deployed on the external cohorts, the model showed a significant correlation with the manual TILs scoring (ρ=0.60, p<0.001). The model yielded an AUC of 0.78 (0.75–0.81) for >5% TILs and AUC of 0.92 (0.88–0.95) for the >50% TILs. Predicted low-TILs tumors were associated with significantly lower 10-year iDFS in TNBC (55.1% versus 70.9%, p=0.034), with a similar pattern observed in HER2-positive (55.3% versus 64.8%, p=0.096), whereas there was no difference in 10-year iDFS between the predicted TILs groups for luminal cancers. Conclusions: These results indicate that Polaris TME accurately quantifies TILs derived from digitized H&E slides and provides prognostic measure for long-term iDFS in high-risk early breast cancer. Prospective validation and harmonization efforts to facilitate clinical implementation are warranted.