Risk score stratification of cutaneous melanomas based on whole slide images using deep learning.

e21573 Background: Predictive and prognostic biomarkers represent the cornerstone of clinical oncology. Improving the prognostication of melanoma is needed to select patients for effective adjuvant therapies. Taking into account that tumor tissue contains a large amount of clinically relevant hidden information that is not fully exploited, we applied a weakly-supervised deep-learning approach on H&E-stained whole-slide image (WSI) to directly predict 5-year survival outcomes in melanoma patients. Methods: We designed a deep neural (DL) network that extracts features from no-padding patches of WSI (tiles of size 512x512 pixels) using a self-supervised learning framework as well as from the tumor region. These features are then fed along with survival information into the deep learning fusion network assigning a survival risk score to each WSI. The model was trained and validated on 195 cases of primary melanoma diagnosed between 2015 and 2017 originating from the IHP Group and the French RICMel melanoma database. Performance was evaluated using a cross-validation and cross-testing framework as well as on an external set of 238 cases from the TCGA database. Concordance index (c-index) was used as a metric to assess the performance of the proposed algorithm. Results: For the prediction of survival, the proposed pipeline yielded c-index of 0.76 and 0.66 for the IHP Group cohort and TCGA dataset respectively. Furthermore, the model is able to significantly discriminate between the 2 groups of patients, with a good and a poor prognosis in terms of overall survival (p-value < 0.001 for IHP and p-value = 0.01 for TCGA). Also, the proposed score is statistically significant when added to a multivariate cox model incorporating the classic prognosis factors (p-value < 0.005) and allows to reach a c-index of 0.81 over the discovery cohort. Conclusions: This infrequent weakly-supervised deep-learning approach to HE images analysis has the advantage of leaving the feature selection entirely to the deep network, and so discover new morphological biomarkers. By this way, it could have in the near future the potential to accelerate and improve the clinical decision-making process in the field of melanoma. Also, these findings will undoubtedly open room for substantial applications with economic and treatment delay fallouts.