Purpose:
To utilize machine learning algorithms to differentiate between malignant and benign hepatic lesions by integrating T2-weighted MRI-based radiomics and apparent diffusion coefficient (ADC) metrics.

Material and methods:
This cross-sectional study (January 2023–December 2024) enrolled 111 patients (71 male, 40 female) aged ≥ 18 years. Hepatic tumors were classified via radiological, pathological, and biochemical assessment. Radiomic features were derived from T2-weighted images, and ADC metrics from diffusion-weighted imaging. The classification performance was evaluated using three models: Model 1 (ADC and ADC Ratios), Model 2 (Radiomic Features), and Model 3 (Combined Features). Stratified 5-fold cross-validation was performed across four algorithms: Logistic Regression (LR), Support Vector Machine (SVM), Gradient Boosting (GB), and Random Forest. The area under the receiver operating characteristic curve (AUC) was used to measure the models’ discriminatory ability. All statistical analyses were conducted using Python libraries.

Results:
A total of 111 hepatic lesions were analyzed: 75 malignant and 36 benign. Among 12 significant radiomic features (p < 0.05), firstorder_InterquartileRange, ngtdm_Contrast, and glszm_ZonePercentage showed the largest effect sizes (Cohen’s d > 0.70). All ADC-related metrics were significantly higher in benign lesions compared to malignant ones (p < 0.001). Model 3 demonstrated the strongest overall performance, with both LR and SVM achieving accuracy and AUC values above 0.800. While Model 1 with SVM offered the highest sensitivity, Model 2 displayed the lowest specificity and weakest overall performance, whereas Model 3 generally showed higher classification performance than either feature set alone.

Conclusions:
Combining T2w-based radiomics features with ADC and ADC ratio values may improve discrimination between malignant and benign liver tumors.