.png)
Breast cancer is the most common cancer among women, with an estimated 2 million new cases and 627,000 deaths globally in 2020. Early diagnosis and appropriate treatment are key to improving patient outcomes and survival rates. Several prognostic biomarkers – clinical, pathological, and molecular – are routinely used to assess the risk of disease progression and the response to targeted therapy, thereby guiding treatment decisions. However, some limitations remain, necessitating new tools to better identify patients at high risk of disease progression and death compared to those at low risk, and to determine the most appropriate therapeutic strategies. In this context, artificial intelligence (AI) applied to digitized histopathological images could serve as a powerful tool to predict disease progression. AI has the potential to detect morphological features not immediately apparent to the naked eye under a microscope. In this study, we aim to develop a deep-learning algorithm to predict the five-year overall survival of patients with breast carcinomas, based solely on the analysis of H&E-stained WSI of breast tumors used for diagnosis.
We introduce a deep neural network (DNN) specifically designed to calculate a survival risk score for patients with breast carcinomas directly from whole slide images (WSI) of H&E-stained sections of the tumor, without requiring any annotations. The model integrates two distinct types of morphological information: one encapsulates cellular-level details, while the other addresses tissue-level features, using the Cox proportional hazard loss function. The model pipeline is presented in Figure 1. The algorithm has been trained and evaluated on the publicly available TCGA-BRCA dataset. Cases lacking a diagnosis date, WSI, or those with marker-pen artifacts were excluded from this study, resulting in a total of 1,003 patients. Among these cases, there were 141 recorded deaths. We employed a 5-fold stratified framework and enhanced both the assessment of our pipeline and the model's robustness through cross-testing and cross-validation techniques. The concordance index (c-index) was used as a metric to assess the performance of the proposed algorithm. Model outcomes were then integrated into a Cox model alongside clinical features such as age and TNM stage to evaluate whether our model could serve as an independent crucial prognostic factor in predicting survival. Additionally, we developed a heatmap on WSI showing Regions Of Interest (ROI) captured by our model to provide the AI-based risk score. An external cohort from the IHP Group of 222 patients was used to further evaluate the model.
The SmartProg model demonstrated an average concordance index of 0.682 when assessed on the testing set. The Cox model, which incorporates clinical features (age, ER/PR, HER2, and TNM stage), produced an average c-index of 0.78. This value increased to 0.79 with the inclusion of our AI-based risk score. These results indicate that the proposed model not only matches but also surpasses existing models in predicting survival, thanks to the robust cross-testing and cross-validation techniques. Additionally, the Cox model suggests that our AI-based risk score can be used as an independent prognostic factor for predicting overall survival (p < 0.005), see Table 1. Furthermore, we were able to significantly discriminate between two groups of patients in terms of survival outcomes, depending on their AI-based high or low risk. A Kaplan-Meier survival curve, provided in Figure 3, combines results from each of the five testing sets and illustrates the stratification of the population into two distinct groups: the high-risk group with a median OS of 8.5 years, and the low-risk group, which has not reached median OS (log-rank test p-value of 9.02e-10). Our findings demonstrate significant potential for forecasting a risk score for breast cancer patients. Finally, an example of a heatmap from a WSI is depicted in Figure 2, clearly showing regions of interest (ROI) in the tumor tissue with important prognostic value for our model. This approach aims to provide an onset of interpretability. The evaluation of the model on an external dataset yielded performance indices of 0.74 and 0.71, with and without the AI-risk score, respectively.
Interestingly, our developed algorithm can automatically extract prognostic morphological features from HE WSI, to predict an AI-based risk score demonstrating a significant prognostic influence in terms of overall survival for breast cancer patients. These results are important as they could aid clinical decision making and improve the quality of care of patients. This emerging and disruptive prognostic approach represents a new concept in the field of precision oncology and personalized medicine. Further studies based on other independent cohorts are required to validate the performance of the algorithm, demonstrate its superiority over the current prognostic markers, as well as to offer further explainability. Indeed, deciphering and understanding the prognostic ROI on tumor tissues will add somes important informations in terms of interpretability of our models, and thus confidence.
J Rynkiewicz, J Paul, Y Salhi, C Bossard, S Salhi, J Chetritt