Automated deep learning for detection and measurement of adrenal masses in contrast-enhanced abdominal CT
- 29 January 2026
- journal article
- urogenital
- Published by Springer Nature in European Radiology
- Vol. 36 (6) , 4527-4538
Abstract
Objectives Adrenal tumors can be functional or malignant, yet they are often overlooked in abdominal CT. This study aimed to develop and validate a fully automated deep learning model for adrenal mass detection and measurement in contrast-enhanced abdominal CT. Materials and methods This retrospective study included 415 CT scans with (n = 155) and without (n = 260) adrenal masses for model development (median age, 40 years; 206 men). Adrenal gland masks were automatically generated using a pretrained segmentation model, and adrenal mass masks were manually refined to train a U-Net–based segmentation network. Two secondary test sets were used for validating mass detection (external test set, n = 995) and size measurement (internal test set 2, n = 50). The external test set reflects a real-world incidence of adrenal masses (4.8%) as determined by radiologic evaluation without pathologic confirmation. The model performance was evaluated using the area under the receiver operating characteristic curve (AUC) for adrenal mass detection and the intraclass correlation coefficient (ICC) for adrenal mass measurement. Results In the internal test set 1 (n = 42), the AUC for classifying adrenal mass was 0.99 (95% confidence interval [CI]: 0.92–1.0), with an average Dice score of 0.812. The model achieved an AUC of 0.94 (95% CI: 0.93–0.96) with sensitivity, specificity, accuracy, and positive predictive value of 89.6%, 96.9%, 96.6%, and 59.7% (95% CI: 77.3–96.5%, 95.6–97.9%, 95.3–97.5%, and 50.6–68.2%, respectively), respectively, in the external test set. The model identified 44 of 50 adrenal masses in internal test 2. The ICCs of the predicted diameter were 0.848 (95% CI: 0.723–0.917) and 0.855 (95% CI: 0.735–0.921) for the CT-measured and pathologically measured diameters, respectively. Conclusion The proposed deep learning model accurately detected and measured adrenal masses in abdominal CT. The model has the potential to improve detection rates of adrenal lesions and facilitate their early management. Clinical relevance statement The proposed deep learning model can detect adrenal masses often missed by radiologists and accurately estimate their size, potentially improving patient management in screening, follow-up, and preoperative settings. Key Points Adrenal masses, which may be functional or malignant, are frequently missed on abdominal CT, leading to delays in appropriate management. A deep learning model can accurately detect the adrenal masses (accuracy 96.6%) and predict the size similar to the sizes measured in CT and surgical specimens (ICC 0.848–0.855). The proposed DL algorithm may help clinicians in various clinical settings, including screening, follow-up, and preoperative evaluation of adrenal masses. Graphical AbstractKeywords
Funding Information
- National Research Foundation of Korea (NRF2022R1F1A107351511)
This publication has 34 references indexed in Scilit:
- Adrenal lesions in patients with abdominal multitraumaEndocrine, 2025
- nnU-Net Revisited: A Call for Rigorous Validation in 3D Medical Image Segmentation2024
- Trustworthy clinical AI solutions: A unified review of uncertainty quantification in Deep Learning models for medical image analysisArtificial Intelligence in Medicine, 2024
- Extending nnU-Net Is All You Need2023
- Fully automatic volume measurement of the adrenal gland on CT using deep learning to classify adrenal hyperplasiaEuropean Radiology, 2022
- Multi-organ Segmentation via Co-training Weight-Averaged Models from Few-Organ Datasets2020
- A novel pipeline for adrenal tumour segmentationComputer Methods and Programs in Biomedicine, 2018
- Multi-Planar Deep Segmentation Networks for Cardiac Substructures from MRI and CT2018
- The Importance of Skip Connections in Biomedical Image Segmentation2016
- Adrenal Gland Abnormality Detection Using Random Forest ClassificationJournal of Digital Imaging, 2013