Survival outcome prediction of breast carcinomas on whole-slide histopathology images using deep learning.

1070 Background: Breast cancer is the most common cancer among women with 2 million new cases and 627,000 deaths in 2020. Early diagnosis and effective treatment are crucial for improved outcomes. Prognosis mainly depends on histopathological features, among them grade. As whole slide histopathology image (WSI) of tumor tissue contains a huge amount of hidden morphological features unexploited by pathologists, we investigate the potential of Artificial Intelligence (AI)-based analysis on WSI to predict prognosis in terms of 5-year overall survival in breast cancer patients. Methods: We used a novel deep neural network (DNN), DiaDeepBreastPRS, designed specifically for predicting a survival risk score in breast cancer patients based on H&E-stained whole slide images (WSI) of tumor tissues. The model incorporates two distinct viewpoints, one capturing cellular details and the other the tissue-level information. DiaDeepBreastPRS was trained and evaluated on a multi-centric discovery cohort of 1,027 patients (1,095 H&E WSI) from the TCGA-BRCA dataset using a cross-testing and cross-validation technique. It was evaluated on an external cohort, comprising 232 patients (247 H&E WSI). A statistical analysis with a multivariate Cox regression model was carried out on clinico-pathological data. AI scores and the concordance index (c-index) serves as the metric for assessing the performance. The predicted risk scores were used for effective risk stratification of breast carcinomas. Results: On the TCGA-BRCA dataset, the model achieved an average c-index of 67%, which reaches 78% by adding the pTNM stage and age at diagnosis. On the external dataset, the model achieved a c-index of 66% and 75% when including some histopathological prognosis factors (pTNM stage, age at diagnosis, HER2 and HR status and mitosis). The AI score was independently associated with the survival of breast cancer patients with the highest hazard ratio (HR 2.46, p<0.005). Furthermore, the model was able to significantly discriminate between the 2 groups of patients, with a good and a poor prognosis in terms of overall survival (p<0.005). Conclusions: In this study, we showcased that the algorithm was able to instantly extract prognostic morphological features from H&E whole slide images (WSI) and could be included in the pathology report. This could potentially enhance clinical decision-making, elevating the standard of care. Compared to commonly used molecular signatures, the AI algorithm enables a reduction in response time and cost savings. However, further investigations using additional independent cohorts are essential to consolidate the algorithm's performance and allow its generalizability, establish its superiority over existing prognostic markers, and provide insights into its interpretability.