Can radiomic feature analysis differentiate adrenal metastases from lipid-poor adenomas on single-phase contrast-enhanced CT abdomen?

ABSTRACT

AIM: To assess if radiomic feature analysis could help to differentiate between the lipid-poor adenomas and metastases to the adrenal glands. MATERIALS AND METHODS: Eighty-six patients (womenmen 4244; mean age 66 years) with biopsy-proven adrenal metastases and 55 patients (womenmen 3916; mean age 67 years) with lipid-poor adenomas who underwent contrast-enhanced, portal-venous phase CT of the abdomen. Radiomic features were extracted using the PyRadiomics extension for 3D Slicer. Following elastic net regularisation, seven of 1,132 extracted radiomic features were selected to build a radiomic signature. This was combined with patient demographics to create a predictive nomogram. The calibration curves in both the training and validation cohorts were assessed using a Hosmer-Lemeshow test. RESULTS: The radiomic signature alone yielded an area under the curve of 91.7% in the training cohort (n=93) and 87.1% in the validation cohort (n=48). The predictive nomogram, which combined age, a previous history of malignancy, and the radiomic signature, had an AUC of 97.2% in the training cohort and 90.4% in the validation cohort. CONCLUSION: The present nomogram has the potential to differentiate between a lipid-poor adrenal adenoma and adrenal metastasis on portal-venous CT.