Prediction of preoperative Ki-67 index in hepatocellular carcinoma patients:a small-sample-based integration of T2WI radiomic features with multiple machine learning algorithms
HUANG Ying
LIU Xuhong
HAN Xiaobing
LIN Tao
HE Guifeng
HUANG Yifeng
DING Bijiao
ZHANG Yonghui
WANG Xinda
ZHANG Qianying
Abstract：Objective To compare the differences in Ki-67 expression levels among newly diagnosed hepatocellular carcinoma(HCC)patients using machine learning methods and to investigate their role in preoperative prediction of immunohistochemical characteristics.Methods This dual-center study collected MRI data and immunohistochemical Ki-67 expression profiles from 59 newly diagnosed HCC patients at the 910th Hospital of Joint Logistics Support Force and Quanzhou First Hospital from November 2023 to October 2024.The cohort included 52 males and 7 females[age 32-55(44.71±6.639)years],stratified into high-expression(Ki-67≥20%,n=37)and low-expression(Ki-67<20%,n=22)groups based on a 20%threshold.All patients underwent preoperative contrast-enhanced MRI(3.0T scanners:Siemens Skyra and GE Discovery 750)with T2WI.Tumor regions of interest were manually delineated using 3D Slicer,and 1,198 radiomic features(shape,first-order statistics,texture,and wavelet transforms)were extracted via the OnekeyAI platform.Synthetic minority oversampling technique addressed class imbalance,followed by rigorous preprocessing(missing value imputation,outlier detection,and data standardization).Eight machine learning models,including logistic regression,support vector machine,K-nearest neighbors,Random forest,extremely randomized trees(ERT),XGBoost,LightGBM,and multilayer perceptron,were implemented for Ki-67 classification.Results Random forest,ERT and XGBoost demonstrated superior performance during training,with XGBoost achieving the highest AUC(0.914),followed by Random forest(0.911)and ERT(0.833).In testing,these models maintained robust generalization capabilities,yielding AUCs of 0.741(Random forest),0.750(ERT),and 0.777(XGBoost),respectively.Conclusion This study demonstrates that a small-sample-based integration of T2WI radiomic features with machine learning algorithms enables effective preoperative prediction of Ki-67 proliferation index in HCC patients.
Keywords：Ki-67liver cancermachine learningROC curvemulti-center study
Publication Date：2025-08-20
Online Publishing Date：2025-09-22(First online date of this platform, not the publication date of the document)
Pages：6( 946-951 )