Abstract 2206: Machine learning radiomics to predict early recurrence of intrahepatic cholangiocarcinoma after curative resection: a multicenter cohort study

Abstract

Abstract Background and Aims: The prognosis of intrahepatic cholangiocarcinoma (ICC) is dismal because of a high incidence of postoperative recurrence. Identifying patients who are likely to experience recurrence is important. We aim to construct radiomics models by machine learning algorithms to predict early recurrence (ER) of ICC after curative resection. Methods: Patients with ICC undergoing curative resection from three institutions were retrospectively recruited and assigned to training and external validation cohorts. Clinical data, contrast-enhanced computed tomography (CECT) images and follow-up information were collected. MRIcroGL and python software were used to segment and extract radiomics features, and student's t-test and mRMR algorithm were performed to further screen features. Logistic regression analysis was used to identify risk factors of ER. Various machine learning algorithms were used to construct predictive models. Predictive performances were evaluated by the receiver operating characteristic (ROC) curves, calibration curves and decision curves analysis (DCA) curves in the external validation cohort. Results: A total of 127 patients with ICC undergoing curative liver resection were enrolled, including 90 patients in the training set and 37 patients in the validation set. Seventy-one (55.9%) patients exhibited ER within one year after surgery, who experienced worse overall survival (OS) than patients with late recurrence and without recurrence (p<0.001).Male gender, vascular invasion, TNM stage III-IV and elevated CA19-9 were independent risk factors of ER, and the resulting clinical model had an AUC value of 0.685.A total of 57 differential radiomics features were identified from pre-treatment CECT images and the 10 most important features were utilized for subsequent modeling. Seven machine learning radiomics models were developed to predict ER of ICC,with the mean AUC value of 0.87±0.02.Among various machine learning algorithms, Random Forest, Neural Network and Support Vector Machine achieved the highest AUCs of 0.89.Additionally, we incorporated clinical data to build seven clinical-radiomics models, of which the Neural Network model had highest AUC value of 0.93.The calibration and DCA curves also indicated that machine learning radiomics models and clinical-radiomics models performed much better than the clinical model. Conclusions: We construct valuable machine learning radiomics models derived from preoperative CECT to predict postoperative ER of ICC, which may serve as an novel non-invasive tool for clinicians to determine surveillance strategies and optimize individual management. Citation Format: Zhiyuan Bo, Bo Chen, Yi Yang, Zhengxiao Zhao, Yi Wang, Gang Chen. Machine learning radiomics to predict early recurrence of intrahepatic cholangiocarcinoma after curative resection: a multicenter cohort study [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 2206.