关键词: 肝细胞癌, 根治性切除术, 肿瘤复发, 机器学习, 多模态数据, 预测

Abstract: Objective This study aimed to build up and validate a machine learning model by integrating clinicopathological data and contrast-enhanced computed tomography (CT) imaging to predict the risk of early tumor recurrence in patients with hepatocellular carcinoma (HCC) with China Liver Cancer (CNLC) stage Ⅰ/Ⅱa after radical tumor resection. Methods 182 patients with CNLC stage I/IIa HCC were encountered in our hospital between July 2020 and June 2023, and all underwent radical hepatectomy. Patients were followed-up for two years after resection. They were randomly divided into a training set (n=127) and a validation set (n=55) at a ratio of 7:3. Clinical and pathological data were collected, and radiomic features were extracted from contrast-enhanced CT scan. Logistic regression analysis, least absolute shrinkage and selection operator (LASSO) regression were applied to screen key features. Clinical and radiomics models were constructed by using adaptive boosting (Adaboost) algorithm and integrating the two types of features to build up a fusion model. The area under the receiver operating characteristic curve (AUC) was drawn and calculated to evaluate the model's risk prediction performance. Results Of the 182 patients with HCC, intrahepatic tumor recurrence were found in 85 cases (46.7%) and not found in 97 cases (53.3%) within 2 years post-surgery; univariate and multivariate Logistic regression analysis showed that serum alpha-fetoprotein levels, tumor diameters and microvascular invasion (MVI) were the risk factors for postoperative tumor recurrence (P<0.05); LASSO regression had selected 8 radiomic features to construct the model (λ_1se=0.085), and by using the Adaboost algorithm to integrate two types of features to construct a fusion model, the AUC to predict in the training set was 0.903 (95% CI: 0.826-0.979), with accuracy (Ac) and specificity (Sp) of 0.856 and 0.859, respectively, while the AUC in the validation set was 0.848 (95% CI: 0.749-0.948), with Ac and Sp of 0.767 and 0.722, respectively; the predictive efficacy in the training set by fusion model was much better than that by clinical model or imaging model(P<0.05) alone. Conclusion The multimodal machine learning model integrating clinical, pathological data and contrast-enhanced CT scan could effectively assess the risk of early tumor recurrence in patients with CNLC stage I/IIa HCC after radical resection, which might provides a reliable tool for individualized postoperative follow-up and intervention strategies.

Key words: Hepatoma, Radical hepatectomy, Tumor recurrence, Machine learning, Multimodal data, prediction