Radiomics-based detection of lateralization in primary aldosteronism: a multicenter study

Abstract Purpose : To investigate whether adrenal CT radiomics can distinguish unilateral from bilateral primary aldosteronism (PA), focusing particularly on inter-lateral asymmetry of radiomic features. Methods: Patients were selected from the multicenter Spanish SPAIN-ALDO registry and included if adrenal computed tomography (CT) imaging and definitive subtype classification by adrenal venous sampling were available. Portal-phase CT scans from 33 patients (66 adrenal glands; 22 unilateral and 11 bilateral PA) were centrally analyzed. Adrenal glands were segmented independently by two radiologists. Radiomic features underwent reproducibility filtering (ICC ≥0.75), redundancy reduction, and bootstrap-based stability selection. Patient-level lateralization was primarily assessed using inter-lateral asymmetry indices, with repeated stratified cross-validation and balanced logistic regression. Results: Two reproducible radiomic features—GLSZM small zone emphasis and GLSZM gray-level non-uniformity—showed consistent discriminatory signals for unilateral versus bilateral PA. Radiomics based on inter-lateral asymmetry achieved an AUC of 0.681, increasing to 0.733 when systolic blood pressure was added to the model. At the individual-gland level, Compactness2 and maximum 3D diameter were selected. Radiomics alone achieved an AUC of 0.660 for identifying the aldosterone-producing gland, with no improvement from adding clinical variables. Morphological features showed excellent interobserver reproducibility. Conclusions: In this multicenter exploratory cohort, reproducible adrenal CT radiomic features demonstrated a potential signal for PA subtype classification. Inter-lateral radiomic asymmetry combined with systolic blood pressure showed the highest point-estimate performance for distinguishing unilateral from bilateral PA, while radiomics alone performed similarly for identifying the aldosterone-producing gland. These findings warrant validation in larger independent cohorts .