Prognostic prediction in soft-tissue sarcomas using deep learning and digital pathology of tumor and margin areas.

The histological FNCLCC grade is the primary prognostic factor in soft-tissue sarcoma (STS) but fails to fully capture high risk patients. This study aimed to develop and validate a deep learning (DL) model to predict metastatic relapse-free survival (MFS) using digital hematoxylin and eosin-stained whole-slide images. A retrospective analysis was conducted on 308 STS patients from two cancer centers, divided into a training cohort (149 patients) and two independent validation cohorts (64 and 95 patients). Supervised multi-instance learning convolutional neural network models were trained on distinct tumor regions-center (C), periphery (P), and margins (R)-to optimize predictive performance. Univariable analysis showed DL models using tumor center (DL-C), periphery (DL-P), and their combination (DL-CP) were consistently associated with MFS across cohorts, while models incorporating margins (DL-R and DL-CPR) demonstrated less reliable associations. Multivariable Cox regression confirmed that high risk scores from DL models were independent predictors of MFS. The DL-CP model outperformed FNCLCC grading in prognostic accuracy, with c-indices ≥ 0.74 in validation cohorts. Adding tumor margin information did not improve predictions.DL models focusing on tumor center and periphery provide superior prognostic value in STS, offering a streamlined, effective approach for digital pathology-based risk stratification.