Radiomics-driven machine learning models for noninvasive prediction of pathological differentiation in hepatocellular
1Department of Radiology, The People's Hospital of Yubei District of Chongqing City, Yubei District, Chongqing, 401120, China.
Aim:
The aim of this study was to compare machine learning and radiomics models across institutions for predicting hepatocellular carcinoma (HCC) pathological grade and evaluate model robustness and generalisability.
Materials And Methods:
This multicentre retrospective study enrolled 321 patients with pathologically confirmed HCC. Radiomics models were developed using pretreatment magnetic resonance imaging (MRI) sequences (arterial phase [AP], T2-weighted imaging [T2WI], diffusion-weighted imaging [DWI], and combined AP + T2WI + DWI). Manual volumetric segmentation was performed for all sequences. Feature selection utilised Least Absolute Shrinkage and Selection Operator (LASSO) regression with five-fold cross validation. Four machine learning classifiers (support vector machine [SVM], random forest [RF], eXtreme Gradient Boosting [XGBoost], and Light Gradient Boosting Machine [LightGBM]) were used. Model performance was evaluated based on discrimination (AUC, accuracy [ACC], recall, precision, and F1-score), clinical utility (decision curve analysis), and calibration (Brier score). Feature importance was interpreted via SHapley Additive exPlanations (SHAP) values.
Results:
A total of 321 HCC patients were included (mean age: 58.55 years ± 11.006 [standard deviation (SD)]; 279 males [86.9%] and 42 females [13.1%]) from multiple medical centres, with histopathologically confirmed diagnosis and pretreatment multiphasic contrast-enhanced MRI examinations. Demographic and clinical characteristics showed no significant intergroup differences (high-grade vs low-grade) except for alpha-fetoprotein (AFP: P=0.004), glypican-3 (GPC-3, P<0.001), and microvascular invasion (MVI, P=0.015). Four machine learning models were developed and comparatively evaluated. The T2WI-/DWI-based RF model demonstrated superior diagnostic performance, while the SVM showed optimal efficacy using AP features. The combined model achieved the highest diagnostic ACC (0.829), outperforming single-sequence radiomics approaches.
Conclusion:
This study developed an MRI radiomics model using machine learning to noninvasively predict the pathological differentiation grade in HCC, with SHAP analysis identifying key discriminative features for preoperative decision support.
