A Machine Learning-based Radiomics Study For The Prediction Of Early Recurrence And Preoperative Peripheral Invasion Status Of Intrahepatic Cholangiocarcinoma

ObjectiveIntrahepatic cholangiocarcinoma(ICC)is a rare and easily misdiagnosed primary liver tumor.It is not only highly malignant and histologically heterogeneous,but also has a poor prognosis with high recurrence and mortality rates after surgery.Finding the risk factors affecting postoperative recurrence of ICC and making accurate preoperative assessment is crucial.However,previous studies were few and the sample size was low(mostly less than 200 cases),which failed to dig deeper into the underlying features and characteristics.In this study,we used multicenter large-sample data and machine learning-based CT radiomics approach to analyze the radiomics characteristics of ICC,predict the risk of early recurrence(ER)and peripheral invasion status,thus providing evidence and data support to optimize the clinical diagnosis and treatment plan of ICC..Methods(1)In the first part of the study,a total of 311 ICC patients from 8 medical centers in China who underwent curative resection were allocated to derivation(n=140),internal validation(n=36)and two external validation(n=74&n=61)cohorts.After generating 6,296 radiomic features from each patient,a machine learning analysis(LightGBM)was used to construct models for ER prediction.The model was validated in multiple centers as well as compared with the existing TNM tumor staging system.In addition,we investigated the interpretability of the LightGBM model.(2)In the second part of the study,220 patients with ICC were retrospectively included and randomly divided into a training group(n=154)and a validation group(n=66)in a ratio of 7:3.After feature selection,quantitative scoring of radiomics features(SVM-RadScore)was performed using support vector machines(SVM)machine learning algorithms.Finally,a Nomogram model was constructed combined with the clinical features to assess the discrimination,precision and clinical benefit value of the model.Results(1)In the first part of the study,a preoperative model that utilized 15 radiomic features and 3 clinical features(CA19-9>1000 U/ml,vascular invasion and tumor margin)was established for evaluating the ER risk,resulting in the area under the curves(AUCs)of 0.974 in the derivation cohort,and 0.871-0.882 in the internal and external validation cohorts,respectively,which are superior to AJCC 8th TNM staging system(AUCs:0.686-0.717).Especially,the sensitivity of this machine learning driven model could reach 94.9%for the entire cohort.(2)In the second part of the study,a Nomogram model was developed by SVM-RadScore(based on six imaging features)and three clinical features(lymph node involvement,tumor vascular invasion and tumor margin),with an AUC of 0.935(95%CI 0.895-0.975)in the training group and 0.843(95%CI 0.7450.941),which was superior to the imaging histology model alone and the clinical features model.The calibration curves suggested a good agreement between the predicted risk of peripheral invasion and the actual observed risk in both the training and validation groups of the Nomogram model.Clinical decision curve analysis suggested that the Nomogram model had better clinical benefit.Conclusions1.The LightGBM machine-learning algorithm-based radiomics model can well identify people at high risk of ER after curative resection of ICC,so that better treatment and follow-up plans can be targeted to reduce the risk of recurrence and prolong survival time,and the model had been validated in multiple centers,confirming its stability and reproducibility;2.The visualized Nomogram model based on CT radiomics features combined with SVM algorithms is a simple and useful tool to predict the risk of peripheral invasion status in patients with ICC preoperatively,helping to accurately assess the disease preoperatively,detect occult peripheral invasion and metastases,and reduce surgical risk,patient burden and loss.