Multimodal prediction of metastatic relapse using federated deep learning in soft-tissue sarcoma with a complex genomic profile

Soft Tissue sarcomas (STS) are a group of heterogeneous and complex diseases where being able to predict the appearance of metastases is key to inform clinical decisions, especially the prescription of adjuvant chemotherapy.

We developed SarcNet: a multimodal deep-learning algorithm based on histological whole-slides and clinical variables to predict metastatic relapse in patients affected with limbs and trunk wall STS.

Two independent series were investigated simultaneously in a privacy-enhanced multicentric setup using Federated Learning: Centre Léon Bérard (n = 221) and Institut Bergonié (n = 390). We then collected two additional validation cohorts from Centre Léon Bérard (n = 93) and Institut Bergonié (n = 124) to validate the performance of the trained algorithm. Evaluated in cross-validation setting on the first two cohorts, SarcNet achieved a performance of 0.797 AUC (Area Under the Receiver-operating characteristic Curve, 0.762–0.833) to predict 5-year MFS comparable to the Sarculator, current state-of-the-art nomogram (0.778 (0.743–0.814)), and outperforming the FNCLCC grading, widely used in practice (0.706 (0.670–0.743), P-value < 0.001).

On validation cohorts, SarcNet performance is still on-par to the Sarculator and FNCLCC rating. Interpretability investigations of SarcNet highlighted histological patterns driving the prediction of metastatic relapse such as atypia, tumor cellularity and anisokaryosis. This model may assist clinical decisions by identifying high-risk patients that could benefit from neoadjuvant or adjuvant treatment.