Prediction of Lymph Node Metastasis from Whole Slide Pathology Images of Early Gastric Cancer using a 2 step Machine Learning Approach

Le résumé fourni par la source

Aims In Early Gastric Cancer(EGC), if the risk of lymph node metastasis(LNM) is negligible, endoscopic curative resection is performed preferentially. If LNM is suspicious, additional gastrectomy with lymphadenectomy should be performed. Recently, as artificial intelligence has revolutionized medical field, several studies have been conducted to predict LNM applying machine learning in stomach. The risk of LNM is determined by primary tumor. It is necessary to pick up on fine and diverse morphologic patterns within H&E stained tissue slide. Therefore, the aim of this study is to develop machine learning algorithm for predicting LNM status in patients with EGC using H&E stained WSI. Methods It is a multicenter study in 5 cohorts including Hanyang university Guri hospital(HGH), Kangbuk Samsung medical center(KBSMC), Seoul ST. Mary’s hospital(SS), International ST Mary’s Hospital(ISH), Korea University medical center(KUMC). The cohort was randomly split into a training and validationas well as test set. Whole slide images were annotated and tessellated into 512*512 pixels using ASAP. The annotation followed areas of tumor tissue using spline on the rectangle ROI by expert pathologists. In the first step, Deep Lab V, SE-ResNext101 to segment differentiated and undifferentiated gastric cancer was trained on 7029 Patches from three datasets of EGC. In the second step, Morphological features inferred by the trained segmentation network from WSI of the primary EGC to predict LNM status. Results The data consisted of 243 cases from 5 cohorts. The number of cases in each cohort was 61, 84, 59, 14, and 25, respectively. The classification results of differentiated/undifferentiated tumor and normal tissue are described We represented differentiated tumors as orange, undifferentiated as brown, and normal cells as light orange(apricot color). When visualizing patch images, ground truth, and predictions from both internal and external datasets, we observed significant similarity. And the performance of the classifier at the patch and slide level, as well as internal and external cohorts. A mean AUROC of 0.7487 and a mean accuracy of 0.7 were achieved in predicting LNM status in the test set using XGBoost. Conclusions Our study is the first proof of concept that machine learning trained with deep learning of whole slide images may be able to predict LNM in early gastric cancer. If this machine learning algorithm is clinically feasible, clinical doctors can determine appropriate treatment to EGC patients right after biopsy. Publication History Article published online: 15 April 2024 © 2024. European Society of Gastrointestinal Endoscopy. All rights reserved. Georg Thieme Verlag KG Rüdigerstraße 14, 70469 Stuttgart, Germany

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