Artificial intelligence may improve the accuracy, consistency, and efficiency of adrenal lesion evaluation, particularly when imaging findings are indeterminate. However, according to a new narrative review, current evidence suggests it should be used to assist physicians, not replace expert interpretation.

Artificial intelligence (AI) techniques, including machine learning (ML), deep learning (DL), and radiomics, have shown promising performance in detecting and characterizing adrenal lesions. But because most studies have been retrospective, conducted at a single center, and lacked external validation, the authors said prospective multicenter studies are still needed before these tools can be used routinely in clinical practice.

The narrative review, published in Endocrine Connections, examined studies published between 2018 and 2025 on AI applications in adrenal imaging. The Warsaw, Poland-based authors searched PubMed, Scopus, and Web of Science and focused on original studies evaluating adrenal gland segmentation, lesion detection, and lesion characterization. Computed tomography (CT) was the most commonly studied imaging modality, accounting for 68% of the studies, followed by magnetic resonance imaging (MRI) and positron emission tomography/computed tomography (PET/CT) at 16% each. Only four studies included patients from multiple centers.

According to the review, adrenal incidentalomas are found in about 5% to 7% of abdominal imaging studies and become more common with age. Although 80% to 85% are benign adenomas, distinguishing them from pheochromocytomas, adrenocortical carcinomas, metastases, and hormonally active tumors can be difficult, especially when imaging findings are inconclusive.

One of the most advanced applications of AI is adrenal gland segmentation or automatically outlining the adrenal glands on imaging. The authors described adrenal gland segmentation as an important step for radiomics analysis, surgical planning, and radiation therapy. Early three-dimensional U-Net models achieved Dice coefficients of 0.69, but newer neural network models reached nearly 0.90, indicating substantially greater accuracy. Recent models also performed well on noncontrast CT scans and in large validation datasets that included more than 2,000 examinations.

AI has also shown promise for detecting adrenal lesions. In one early study, an AI model identified adrenal lesions with 92% sensitivity and 91% specificity. In another study, combining AI with radiologist interpretation increased sensitivity to nearly 100% and specificity to 99%. However, the reviewers noted that evidence remains limited for detecting small incidental adrenal lesions in diverse real-world patient populations.

Several studies evaluated whether radiomics could help identify hormonally active adrenal adenomas before biochemical testing. In one study of 206 patients, a CT-based radiomics model predicted mild autonomous cortisol secretion with 81% specificity in the validation cohort. Across multiple studies, models distinguishing hormonally active from nonfunctioning adenomas achieved area under the curve (AUC) values ranging from about 0.80 to 0.99. Performance was highest when imaging features were combined with clinical information rather than imaging alone.

Distinguishing benign from malignant adrenal lesions was another major focus. Lipid-poor adenomas are especially challenging because they often resemble malignant tumors on conventional imaging. In studies comparing lipid-poor adenomas with pheochromocytomas, machine learning models achieved AUC values of about 0.92 to 0.99, with several models maintaining high performance in external validation. One explainable AI model identified a maximum pixel attenuation greater than 125 Hounsfield units as a strong predictor of pheochromocytoma and achieved an F1 score of 0.93. Because the model showed which imaging feature drove its prediction, it may help physicians better understand and trust the results.

AI also showed strong performance in distinguishing adrenal metastases from benign lesions in patients with cancer. One PET/CT model that combined imaging and clinical information achieved an AUC of 0.94. Another model that relied on imaging alone achieved an AUC of 0.91 with 100% specificity. The review suggested these approaches could eventually reduce unnecessary biopsies if future prospective studies confirm their performance.

AI also showed strong performance in identifying adrenocortical carcinoma. In one study, a machine learning model distinguished adenomas larger than 4 centimeters from adrenocortical carcinoma with 82% accuracy, compared with 69% for experienced radiologists. Another deep learning model distinguished nonmetastatic adrenocortical carcinoma from lipid-poor adenomas with sensitivities ranging from 96% to 100% and accuracies of 87% to 91%, depending on how the model was optimized.

The review also highlighted AI systems that classify adrenal lesions into multiple diagnostic categories. Across several studies, these models achieved accuracies greater than 80%, with some CT-based models exceeding 95% for accuracy, sensitivity, and specificity.

For clinicians, the review suggests that the greatest near-term value of AI may be in helping characterize adrenal incidentalomas with indeterminate imaging findings rather than replacing established diagnostic approaches. Models that combine imaging with clinical and biochemical information consistently outperformed imaging-only systems, supporting the development of multimodal decision-support tools that integrate into existing radiology workflows and picture archiving and communication systems.

The review also highlighted several barriers to implementing AI in clinical practice. Most published studies were retrospective, included relatively small numbers of patients, and lacked external validation. Differences in imaging protocols across institutions also made results less reproducible. Model performance often declined when applied to images from different scanners or different patient populations, particularly when models trained on adults were used in children. Other challenges included inconsistent electronic health record systems, limited data sharing, regulatory hurdles, implementation costs, and the need for explainable AI models before routine clinical use.

“AI has strong potential to improve adrenal imaging by enabling faster, more accurate, and reproducible lesion assessment,” the review’s authors concluded. “However, clinical implementation requires prospective validation, explainable models, and standardized integration.”

The work was funded by grants from the Polish Ministry of Science and Higher Education and the INTEGRA 2 project funded by the Medical University of Warsaw and Warsaw University of Technology. The authors reported no conflicts of interest.