Thursday, December 4 | 2:20 p.m.-2:30 p.m. | SSGU07-6 | Room E353B
Researchers have developed an automated machine-learning tool that they say can detect adrenal masses at scale in routine clinical practice.
To be presented by Borna Dabiri, MD, PhD, from Massachusetts General Brigham, the model leveraged 68,213 consecutive scans performed in 42,291 patients, of which 23,658 were female (mean age, 60.9 years). Basic demographic data was extracted from the electronic health record for the Institutional Review Board-approved retrospective study.
Patients underwent unenhanced and contrast-enhanced abdominal CT examinations from June 1, 2023, through March 31, 2025. The model classified adrenal glands as normal or containing a mass. As a reference standard, radiology reports from these exams were categorized as normal or containing an adrenal mass > 1 cm using the Qwen2.5 large language model (LLM), according to the group.
Prior to the RSNA, Dabiri and colleagues highlighted an overall accuracy of the segmentation model -- using the report as ground truth -- at 92.5% with a sensitivity of 55.2% and a specificity of 95.5%.
Ultimately, the group concluded that the automated machine-learning adrenal segmentation and classification model was able to detect adrenal masses with high accuracy and specificity but lower sensitivity in a large population of consecutive CT examinations.
Attend this talk on Thursday afternoon to get more information.
