Published online Dec 28, 2025. doi: 10.4329/wjr.v17.i12.112911
Revised: September 15, 2025
Accepted: December 3, 2025
Published online: December 28, 2025
Processing time: 138 Days and 22.4 Hours
Despite the promising prospects of using artificial intelligence and machine learning (ML) for disease classification and prediction purposes, the complexity and lack of explainability of this method make it difficult to apply the constructed models in clinical practice. We developed and validated an interpretable ML model based on magnetic resonance imaging (MRI) radiomics and clinical fea
To develop and validate an interpretable ML model for preoperative pathological grade prediction in HCC patients via a combination of multisequence MRI radiomics and clinical features.
MRI and clinical data derived from 125 patients with HCCs confirmed by postoperative pathological examinations were retrospectively analyzed. The patients were randomly split into training and validation groups (7:3 ratio). Univariate and multivariate logistic regression analyses were performed to identify independent clinical predictors. The tumor lesions observed on axial fat-suppressed T2-weighted imaging (FS-T2WI), arterial phase (AP), and portal venous phase (PVP) images were delineated in a slice-by-slice manner using 3D-slicer to generate volumetric regions of interest, and radiomic features were extracted. Interclass correlation coefficients were calculated, and least absolute selection and shrinkage operator regression were conducted for feature selection purposes. Six predictive models were subsequently developed for pathological grade prediction: FS-T2WI, AP, PVP, integrated radiomics, clinical, and combined radiomics-clinical (RC) models. The effectiveness of these models was assessed by calculating their area under the receiver operating characteristic curve (AUC) values. The clinical applicability of the models was evaluated via decision curve analysis. Finally, the contributions of the different features contained in the model with optimal performance were interpreted via a SHapley Additive exPlanations analysis.
Among the 125 patients, 87 were assigned to the training group, and 38 were assigned to the validation group. The maximum tumor diameter, hepatitis B virus status, and monocyte count were identified as independent predictors of pathological grade. Twelve optimal radiomic features were ultimately selected. The AUC values obtained for the FS-T2WI model, AP model, PVP model, radiomics model, clinical model, and combined RC model in the training group were 0.761 [95% confidence interval (CI): 0.562-0.857], 0.870 (95%CI: 0.714-0.918), 0.868 (95%CI: 0.714-0.959), 0.917(95%CI: 0.857-0.959), 0.869 (95%CI: 0.643-0.973), and 0.941 (95%CI: 0.857-0.945), respectively; in the validation group, the AUC values were 0.724 (95%CI: 0.625-0.833), 0.802 (95%CI: 0.686-1.000), 0.797 (95%CI: 0.688-1.000), 0.901(95%CI: 0.833-0.906), 0.865 (95%CI: 0.594-1.000), and 0.932 (95%CI: 0.812-1.000), respectively. The combined RC model demonstrated the best performance. Additionally, the decision curve analysis revealed that the combined RC model had satisfactory prediction efficiency, and the SHapley Additive exPlanations value analysis revealed that the “FS-T2WI-wavelet-HLL_gldm_Large Dependence High Gray Level Emphasis” feature con
An interpretable ML model based on MRI radiomics provides a noninvasive tool for predicting the pathological grade of HCCs, which will help clinicians develop personalized treatment plans.
Core Tip: Despite the promising prospects of using artificial intelligence and machine learning for disease classification and prediction purposes, the complexity and lack of explainability of these methods make it difficult to apply the constructed models in clinical practice. This study aimed to develop and validate an interpretable machine learning model for conducting preoperative pathological grade prediction in hepatocellular carcinoma patients via a combination of multisequence magnetic resonance imaging radiomics and clinical features, which will help clinicians better understand the situation and develop personalized treatment plans.