An integrated radiopathomics machine learning model to predict pathological response to preoperative chemotherapy in gastric cancer

Yaolin Song,Shunli Liu,Xinyu Liu,Huiqing Jia,Hailei Shi,Xianglan Liu,Dapeng Hao,Hexiang Wang,Xiaoming Xing

DOI: https://doi.org/10.1016/j.acra.2024.08.014

2024-08-29

Abstract:Rationale and objectives: Accurately predicting the pathological response to chemotherapy before treatment is important for selecting the appropriate treatment groups, formulating individualized treatment plans, and improving the survival rates of patients with gastric cancer (GC). Methods: We retrospectively enrolled 151 patients diagnosed with GC who underwent preoperative chemotherapy and surgical resection at the Affiliated Hospital of Qingdao University between January 2015 and June 2023. Both pretreatment-enhanced computer technology images and whole slide images of pathological hematoxylin and eosin-stained sections were available for each patient. The image features were extracted and used to construct an ensemble radiopathomics machine learning model. In addition, a nomogram was developed by combining the imaging features and clinical characteristics. Results: In total, 962 radiomics and 999 pathomics signatures were extracted from 106 patients in the training cohort. A fusion radiopathomics model was constructed using 13 radiomics and 5 pathomics signatures. The fusion model showed favorable performance compared to single-omics models, with an area under the curve (AUC) of 0.789 in the validation cohort. Moreover, a combined radiopathomics nomogram (RPN) was developed based on radiopathomics features and the Borrmann type, which is a classification method for advanced GC according to tumor growth pattern and gross morphology. The RPN showed superior predictive performance in the training (AUC 0.880) and validation cohorts (AUC 0.797). The decision curve analysis showed that RPN could provide favorable clinical benefits to patients with GC. Conclusions: RPN was able to predict the pathological response to preoperative chemotherapy with high accuracy, and therefore provides a novel tool for personalized treatment of GC.