Abstract B043: Cellular localization as a prognostic indicator in invasive breast cancer

Abstract Background: Breast cancer is the most common cancer type in women, and outcomes can vary widely depending on tumor aggressiveness. Hence, researchers have sought to characterize a patient’s risk to ensure those with more dangerous lesions get more aggressive therapies. In histopathology this is traditionally done using morphological features such as tubule formation and nuclear pleomorphism. Recent advancements in machine learning and scanner quality have allowed for more precise quantification of individual cells. We aim to use these cell densities and their spatial organization within the tumor microenvironment to place patient risk. Methods: Using publicly available data from 4 separate sources, we trained A) a model to segment lymphocytes, and tissue into three classes: tumor, tumor associated stroma (TAS), and other. B) Mitosis detection. C) Red blood cell segmentation. D) Fibroblast segmentation. For the fibroblasts, public data was unavailable, so we used Hovernet to generate a labelled dataset of fibroblasts from TCGA-BRCA, which was used to train a faster model for whole slide segmentation. Subsequently, these cell densities were measured across our tissue types to see their relative concentrations across the whole slide image, providing context on the spatial distribution. We standardized these features and combined them with the hormone receptor statuses (ER, PR, and HER2) in an ElasticNet-regularized cox proportional hazard model trained on 714 images from the Ontario Tumour Board. This yielded 8 factors with non-zero coefficients which were in descending order of magnitude: PR status, lymphocyte-TAS density, tumor density, fibroblast-TAS density, lymphocyte-other density, mitosis-tumor density, red blood cell-TAS density, and red blood cell-tumor density. We also trained a second model using only the imaging features and ignoring PR status to compare to state-of-the-art end-to-end risk prediction methods. Results: To evaluate the generalizability of our model, we used TCGA-BRCA cohort as a test set. Here using the patient’s progression free interval time as an outcome variable, our imaging only model had a c-index of 0.651. This compares favorably to end-to-end multiple instance AI models such as DGNN which had a much lower c-index of 0.570. We believe that this represents an inability for the end-to-end models to map complex relationships between cell types and their localization. Another point of comparison is the histomic prognostic signature (HiPS). This model similarly used cell segmentation to describe slide-wide features. They however focus more on cell-cell relationships and appearance characteristics while we focus more heavily on localization. Both models incorporated hormone receptor status information. The HiPS model had a c-index of 0.563, while our model when given the hormone receptor status was able to attain a c-index of 0.717. Conclusion: We believe this shows that an approach focusing on the interaction between cell types and their location within the tumor microenvironment can produce a robust risk score. Citation Format: Matthew JM. McNeil, Vishwesh Ramanathan, Lincoln Stein, Dianne Chadwick, Anne L. Martel. Cellular localization as a prognostic indicator in invasive breast cancer [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr B043.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.