MRCP-derived radiomics features associated with biliary dilation: a multi-step feature reduction, multivariable...

To identify MRCP-derived radiomics features associated with image-defined biliary dilation, evaluate their independence from demographic confounders, and assess diagnostic classification performance, including in borderline-caliber cases.

In this retrospective cross-sectional study, 135 patients (dilated: n = 60; non-dilated: n = 75) who underwent clinically indicated MRCP (September 2025-January 2026) were included. Two blinded radiologists independently classified biliary dilation (Cohen's kappa = 0.87). Three-dimensional CBD segmentation was performed in 3D Slicer (v5.10.0), with feature extraction using PyRadiomics in compliance with IBSI guidelines. Feature reduction applied sequential reproducibility filtering (ICC (2,1) ≥ 0.80), low-variance filtering, Spearman correlation filtering (|r| > 0.90), Benjamini-Hochberg FDR correction (q < 0.05), and Cliff's delta thresholding (|δ| > 0.5). Multivariable linear regression adjusted for age, sex, and post-cholecystectomy status. Shape features were excluded from classification analyses to avoid incorporation bias. Three classifiers (Random Forest, Logistic Regression, and Naïve Bayes) were trained on the seven Bonferroni-confirmed texture features using five-fold stratified cross-validation. A post hoc exploratory analysis in borderline CBD caliber (6-10 mm; n = 45) used a two-feature model with internal tenfold cross-validation.

Twelve features met selection thresholds: four shape and eight texture features. The strongest discriminators were RunLengthNonUniformity (δ = 0.870; q = 1.09 × 10^-16) and GrayLevelNonUniformity (δ = 0.861; q = 1.43 × 10^-16). Multivariable regression demonstrated that 11 of 12 features were significant independent correlates of image-defined biliary dilation after Bonferroni correction (all p < 0.001). In the full cohort (n = 135; seven texture features), Random Forest achieved AUC = 0.963, accuracy = 92.1%, sensitivity = 94.3%, and specificity = 89.5%. In the exploratory borderline subgroup (n = 45; 6-10 mm; two-feature model), Random Forest achieved AUC = 0.832, suggesting potential discriminatory utility in equivocal cases.

MRCP-derived radiomic features are associated with image-defined CBD dilation and show exploratory internal cross-validated diagnostic performance, including in borderline cases, using standard clinical acquisitions. Prospective multicenter validation with independent clinical reference standards is required before clinical implementation.