Machine Learning-Based Integration and Comparison of ADC Map Radiomics with Conventional Imaging Markers for Cholesteatoma Diagnosis

Abstract

Purpose To compare the diagnostic performance of apparent diffusion coefficient (ADC) map-based radiomics with conventional CT and DWI for differentiating cholesteatoma from non-cholesteatomatous middle ear lesions and to evaluate the incremental value of their integration using machine learning (ML). Methods This retrospective single-centre study included a derivation cohort of 51 histopathologically confirmed lesions (23 cholesteatomas, 28 non-cholesteatomatous lesions) and a held-out internal validation cohort of 20 lesions (9 cholesteatomas, 11 non-cholesteatomatous lesions). Conventional imaging variables included CT attenuation, visually assessed diffusion restriction on DWI, and quantitative ADC. ADC maps were manually segmented, and 122 radiomic features were extracted. Predictive modelling used 3-fold cross-validation with fold-specific feature screening, normalisation, feature selection, and classifier selection. Final combined-model reproducibility was assessed using intraclass correlation coefficients. Results Compared with non-cholesteatomatous lesions, cholesteatomas showed higher CT attenuation, more frequent diffusion restriction, and lower ADC values (allp < 0.05). The final combined model incorporated quantitative ADC, GLRLM low gray-level run emphasis, and first-order coefficient of variation. Among the evaluated classifiers, a three-layer artificial neural network performed best for the combined dataset. In the held-out internal validation cohort, the combined radiologyradiomics model showed the highest diagnostic performance, with 95.0% accuracy, 100% sensitivity, 90.9% specificity, and anAUC of 0.951. Conclusion ADC map-based radiomics may provide complementary microstructural information beyond conventional imaging. In this internally validated exploratory study, combining quantitative ADC with selected ADC-derived radiomic features improved discrimination of cholesteatoma from non-cholesteatomatous middle ear lesions, although external validation is required before clinical implementation.