Deep learning for predicting lymph node metastasis in T1b gastric cancer using histopathology: a retrospective, multicenter study

Le résumé fourni par la source

Background: T1b gastric cancer (T1b-GC) carries a high risk of lymph node metastasis (LNM), which is a key determinant of treatment choice and prognosis. However, assessing LNM risk in T1b-GC remains challenging. This study aimed to develop machine learning models that integrate clinical data and postoperative hematoxylin and eosin-stained whole-slide images (HE-WSIs) to predict LNM in T1b-GC. Methods: This retrospective, multicenter cohort study analyzed 1023 patients with T1b-GC who underwent radical gastrectomy with lymphadenectomy. Thirteen machine learning algorithms were evaluated to develop a baseline predictive model using clinical variables. A deep learning system (DL-T1b) was trained and validated to predict LNM from histopathological tumor sections. A nomogram combining the baseline model and DL-T1b score was developed and evaluated. Results: Linear discriminant analysis was the optimal baseline model, achieving an area under the curve (AUC) of 0.771 (95% CI: 0.747, 0.813) based on six factors: sex, tumor size, differentiation grade, Laurén classification, tumor location, and lymphovascular invasion. The DL-T1b model demonstrated superior performance (AUC: 0.910; 95% CI: 0.869, 0.945). Incorporating the DL-T1b score with baseline clinical features into a nomogram enhanced its discriminatory performance (AUC: 0.964; 95% CI: 0.949, 0.978), achieving 93.2% sensitivity and 85.1% specificity. Spatial heatmaps of HE-WSIs revealed that the proportion of patches with a 100% predicted probability correlated positively with LNM risk. Conclusion: Derived from HE-WSI tumor regions, the DL-T1b score provides a robust and interpretable tool for individualized LNM risk stratification to guide clinical treatment.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.