Deep learning can accurately predict the prognosis of gynecologic smooth muscle tumors of uncertain malignant potential: a multicenter pilot study.

Smooth muscle tumors of uncertain malignant potential of the gynecologic tract (STUMP) are a heterogeneous group of tumors, with ambiguous or worrisome features, whose biological behavior is difficult to predict. Several ancillary techniques have been used to try to predict their prognosis, with limited success. The aim of this study is to explore whether deep learning (DL) based features can be used to predict progression-free survival (PFS) in STUMP and identify high-risk patients, directly from histological slides. A cohort of 95 STUMP was collected from 7 academic centers (79 for training and 16 for external validation). Non overlapping tiles were extracted from the tumor area and used to train a DL model to predict PFS. Python's scikit-learn library and the R software environment were used for data analysis. After 4-fold cross-validation, mean C-indexes of 0.7052 (95%CI: 0.4951-0.9152) and 1.0 (95%CI: 1.0-1.0) were achieved, in the training and external validation cohorts, respectively. The predicted PFS probabilities were used to classify the patients into low-risk and high-risk groups, based on the thresholds of the median and the first quartile of predicted PFS probabilities. Significant differences between both groups were observed, at 10 years, with both thresholds. Cox regression analysis showed that the output of the DL model was associated with a worse prognosis (p = 0.0356). Both STUMP groups were compared with a cohort of leiomyomas (n = 160) and leiomyosarcomas (n = 58). The lowest hazard ratio was observed in leiomyomas, followed, consecutively, by low-risk STUMP, high-risk STUMP and leiomyosarcomas. The Cox model showed good discriminatory potential between the four groups (all pairwise comparisons were statistically significant). These findings suggest that DL-based features can be used for outcome prediction of STUMP. Additional work is needed to establish whether this "high-risk" group can be identified via molecular markers and used to tailor patient surveillance.