The predictive potential of the spatial signature of lymphocytes in breast cancer patients.

Abstract

530 Background: Tumor-infiltrating lymphocytes in breast cancer have emerged as both a prognostic and a potentially predictive immunotherapy biomarker. Advancements in artificial intelligence can extract pathology-based spatial immune fingerprints for use as treatment decision support tools. Methods: We examined 908 primary breast cancer patients with whole slide images (WSI) available from TCGA database. Digital structuring of WSIs included automated detection of lymphocytes, tumor and tumor adjacent stroma, using deep learning-based semantic segmentation. Prognosis was defined as progression free interval (PFI). A Cox Survival analysis was used to detect prognostic spatial features. We used principal component analysis (PCA) to reduce and decorrelate significant features. The resulting PCA features were used to fit the final model. The model was then validated on an independent database of WSI of breast lumpectomies, from two tertiary hospitals in Israel. Results: The analysis included 908 WSI. The average age was 58.4 years old, with a majority of early stage breast cancer (76.7%, stage I and II). The detection performance for tumor area and lymphocytes reached F1 scores of 99% and 97% respectively, in comparison to human annotation. In the Kaplan Meier (KM) analysis of 414 early stage luminal breast cancers, a high number of lymphocyte clusters (LC) and a high ratio between stromal lymphocyte density and tumor lymphocyte density (LD-S/LD-T) were significantly associated with longer PFI (p = 0.005 and p = 0.038, respectively). Based on these features, two continuous PCA features were added to the multivariate model, and remained significantly associated with PFI after adjusting for age (HR = 1.19, 95% CI 1.05-1.35; HR = 1.26 95% CI 1.03-1.55). The validation set was underpowered (n = 79) and data is still being collected. In a preliminary KM analysis of 37 early stage luminal breast cancer cases from the validation set, LD-S/LD-T was significantly associated with longer PFI (p = 0.046). Conclusions: In our study, LC and LD-S/LD-T, presumably surrogate measures of peritumoral lymphocytes, were found significantly associated with longer PFI.