Development and validation of a machine learning-based radiomics model using 45-keV virtual monoenergetic images for differentiating fat-poor angiomyolipoma from clear cell renal cell carcinoma.

Authors

Abstract

Preoperative differentiation between fat-poor angiomyolipoma (fp-AML) and clear cell renal cell carcinoma (ccRCC) remains challenging because of overlapping enhancement patterns on conventional imaging, which may lead to unnecessary surgery. However, the diagnostic value of radiomics derived from 45-keV virtual monoenergetic images (VMIs) for this task remains insufficiently explored. This study aimed to evaluate the diagnostic performance of a radiomics-based approach using 45-keV VMIs from dual-layer spectral detector computed tomography (DLCT) for differentiating fp-AML from ccRCC. In this retrospective study, 108 consecutive patients who underwent contrast-enhanced DLCT urography were enrolled, including 50 patients with pathologically confirmed fp-AML and 58 patients with pathologically confirmed ccRCC. Only small renal masses (≤4 cm) without visible fat on unenhanced computed tomography (CT) were included. Patients were randomly allocated to training (n=75) and validation (n=33) cohorts at a 7:3 ratio. Volumes of interest (VOIs) were manually segmented on 45-keV VMIs from both corticomedullary phase (CMP) and nephrographic phase (NP) scans. After radiomics feature extraction, feature selection was performed using minimum redundancy maximum relevance (mRMR) followed by least absolute shrinkage and selection operator (LASSO). Five machine-learning classifiers were trained and compared using receiver operating characteristic (ROC) analysis, and model interpretability was evaluated using Shapley additive explanations (SHAP). Significant differences were observed between fp-AML and ccRCC in sex, maximum lesion diameter, unenhanced CT attenuation, pseudocapsule sign, and cystic changes. Across the five evaluated classifiers, the dual-phase (CMP + NP) radiomics models showed consistently higher diagnostic performance than the single-phase models. The support vector machine (SVM) with the dual-phase protocol achieved the best performance in the validation cohort, with an area under the curve (AUC) of 0.915 [95% confidence interval (CI): 0.819-1.000], accuracy of 84.8%, sensitivity of 86.7%, and specificity of 83.3%. The CMP and NP models showed lower validation AUCs of 0.844 (95% CI: 0.702-0.987) and 0.756 (95% CI: 0.587-0.924), respectively. A radiomics-based machine learning model using 45-keV VMIs demonstrated good diagnostic performance for differentiating fp-AML from ccRCC in small renal masses. Nevertheless, owing to the retrospective single-center design and the lack of external validation, these findings should be interpreted cautiously, and further multicenter studies are warranted to validate their generalizability and clinical utility.

Tags

Topics