Endocrine Abstracts

Background: Adrenal lipid poor adenoma (LPA) and adrenocortical cancer (ACC) may overlap in computerized tomography (CT). Radiomics recently emerged as new tool for malignant behavior identification.

Aim: To assess radiomics utility for identification of ACC and LPA in adrenocortical masses with unenhanced (UE) CT scan attenuation≥10 Hounsfield Unit (HU).

Methods: We retrospectively enrolled 50 patients, 38 radiologically defined LPA with 6-12 months of radiologic stability or benign histological exam (n=11), and 12 ACC with histological exam (2 patients with Weiss score=3; 4 patient with ki67≥10%). All patients underwent CT with UE scan, arterial (ACE), venous (VCE) and 15’ delayed (DCE) contrast enhanced phases, on which radiomics was performed with LIFEx software (©LITO 2022-2023). We performed a two-steps multivariate analysis for each CT phase to evaluate predictors of malignancy (Weiss score≥3). Multivariate analysis first step was completed within single radiomics feature classes, then first step predictors were altogether employed for multivariate analysis second step. Second step predictors were utilized for receiver operating characteristic curve analysis and estimation of positive (PPV) and negative predictive value (NPV). We conducted multivariate analysis for each CT phase, within single radiomics feature classes, to evaluate predictors of ki67 values, as aggressivity marker.

Results: In UE, surface to volume ratio (SVR) and Run Length Non-Uniformity (RLNU) predicted malignancy (Odds Ratio (OR)=2.718; 95% Confidence Interval (CI)=1.56-4.75; P<0.001), with 83.3% sensibility, 94.3% specificity, 83.3% PPV, 94.7% NPV. Model including HU total lesion glycolysis (TLG), standard deviation, variation coefficient and maximum gray level predicted ki67 (R²=0.984; P<0.001). In ACE, SVR and Feret diameter predicted malignancy [OR=2.718; 95% CI=1.57-4.745; P<0.001], with 83.3% sensibility, 92.1% specificity, 76.9% PPV, 94.6% NPV. Model including HU TLG, energy, variation coefficient and total calcium score predicted ki67 (R²=0.998; P<0.001). In VCE, SVR and compacity predicted malignancy [OR=2.719; 95% CI=1.54-4.79; P<0.001], with 83.3% sensibility, 92.1% specificity, 76.9% PPV, 94.5% NPV. Model including HU maximum histogram gradient, root mean square and intensity histogram minimum grey level predicted ki67 (R²=0.979; P<0.001). In DCE, SVR and RLNU predicted malignancy [OR=2.718; 95% CI=1.54-4.79; P<0.001], with 83.3% sensibility, 91.9% specificity, 76.9% PPV, 94.5% NPV. Model including HU TLG and median absolute deviation predicted ki67 (R²=0.98; P<0.001).

Conclusion: Radiomics seems useful to identify adrenal masses nature, even without CT contrast enhanced phases. Among other radiomics parameters, SVR and HU intensity-based features seem to be powerful predictors of adrenocortical masses malignancy and aggressivity.