Deep learning algorithms from histopathological images stratify molecular subtypes for leiomyosarcoma: a proof-and-concept diagnostic study
Shaohui He, Jun Chen, Yan Fang Liu, Yanbin Xiao, Mei Li, Qi Zhang, Dongjie Jiang, Mengchen Yin, Xin Jiang, Na Cui, Zhengwei Zhang, Wei Wei, Shangjiang Yu, Yao Zhang, Xiaopan Cai, Haifeng Wei, Jianru Xiao · International Journal of Surgery · 2025
BACKGROUND: The leiomyosarcoma (LMS) is the most common soft tissue sarcoma, and its molecular subtypes were identified with therapeutic sensitivity and prognosis. We aimed to develop and validate deep learning (DL) algorithms to stratify molecular subtypes and predict survival by using single hematoxylin-eosin stained whole slide images (WSIs). METHODS: The DL models were trained on the single WSIs ( n = 154, tiles = 1 579 215) from The Cancer Genome Atlas, and externally tested in a multicenter cohort from real world ( n = 80, tiles = 555 211). The primary performance metric was area under the receiver operating characteristic curve (AUROC), others included accuracy, recall, specificity, precision, and F1 score. The computation visualizations (CVs) were further performed to visualize the histomorphological features, and the effect was evaluated on assisting pathologists in subtyping. RESULTS: After five-fold cross-validation, the LMS_DL model based on DenseNet121 achieved an AUROC of 0.944 ± 0.001 in molecular subtyping, while the ResNet50-based DL algorithm achieved highest AUROC of 0.937 ± 0.024 in predicting 2-year overall survival. The LMS_DL model outperformed pathologists by over 30% accuracy in subtyping ( P < 0.001). The histomorphological features summarized by CVs enabled pathologists to obtain accuracy improvements in subtyping by 12.1% ± 4.4% ( P = 0.024) with less diagnostic time ( P = 0.025) and uncertainty ( P = 0.007). CONCLUSIONS: The LMS_DL models can be favorably applied in the molecular subtyping and survival prediction for LMSs to greatly alleviate the workload of pathologists with high accuracy and efficacy, which requires large prospective cohort for further validation.