The application of artificial intelligence in adrenal imaging: current state of knowledge, challenges, and future directions.
Authors
Affiliations (3)
Affiliations (3)
- Student Scientific Club "Endocrinus" Affiliated to the Department of Internal Medicine and Endocrinology, Medical University of Warsaw, Warsaw, Poland.
- Department of Internal Medicine and Endocrinology, Medical University of Warsaw, Warsaw, Poland.
- Faculty of Electrical Engineering, Warsaw University of Technology, Warsaw, Poland.
Abstract
Adrenal incidentalomas are frequently detected on abdominal imaging and require evaluation for malignancy and hormonal activity. Although most are benign adenomas, distinguishing them from pheochromocytomas, adrenocortical carcinomas, metastases, and functioning tumors remains difficult, especially in indeterminate cases. Artificial intelligence as emerged as a promising tool to improve adrenal lesion detection, segmentation, and characterization. This narrative review was based on a structured search of PubMed, Scopus, and Web of Science for studies published between 2018 and 2025 using the terms "artificial intelligence," "machine learning," "deep learning," "radiomics," "adrenal glands," and "adrenal imaging." Original studies on segmentation, detection, and lesion characterization were prioritized. CT was the predominant imaging modality, followed by MRI and PET/CT. AI applications in adrenal imaging include gland segmentation, lesion detection, functional assessment, and lesion classification. Detection models showed high sensitivity and specificity, although evidence for small incidentalomas in heterogeneous real-world datasets remains limited. For lesion characterization, radiomics-and deep learning-based models demonstrated promising performance in differentiating functioning from non-functioning adenomas, lipidpoor adenomas from pheochromocytomas, benign lesions from metastases, and adenomas from adrenocortical carcinoma. Multimodal models combining imaging and clinical data often outperformed imaging-only approaches. Despite encouraging results, most studies were retrospective, single-center, and based on small or selected cohorts, limiting reproducibility and generalizability. Variability in imaging protocols, lack of external validation, and limited workflow integration remain key barriers. At present, AI is best viewed as a decision-support tool. Broader clinical implementation will require prospective multicenter validation, methodological standardization, and more explainable models.