Personalized adrenal gland volume reference ranges and development of a fully automated deep learning screening tool.

OBJECTIVE: This study aims to develop a low-dose CT-based, fully automated deep learning tool for screening adrenal gland volume abnormalities and establishing personalized reference ranges to assist in diagnosing adrenal diseases. METHODS: This study included subjects (≥18 years) who underwent low-dose non-contrast chest CT during routine health check-ups, and three datasets were extracted based on specific criteria: healthy reference, hypertension/diabetes validation, and adrenal abnormality validation. Randomly sampled from these datasets 400 low-dose chest CT images were used to train the nnU-Net-based deep learning model, and 550 images were used for validation. Multivariable regression and restricted cubic splines (RCS) assessed the effects of factors like age, sex, body surface area, and blood markers on adrenal volume. A quantile regression model was used to create individualized reference ranges. A GMM-based anomaly detection system was developed for abnormal volume screening, tested on datasets for hypertension, diabetes, and adrenal hyperplasia. RESULTS: Among the 18,538 adults, 7,907 (42.65 %) were healthy. Adrenal volume ranges ere 3.24 cm3 (2.49-4.06) on the right and 4.51 cm3 (3.72-5.39) on the left. Adrenal volume correlated with age, sex, blood pressure, and glucose (p < 0.05). The parameter-adjusted reference ranges were defined as: Lower limit = -2.739 + 0.016 × Age + 0.431 × Sex + 0.032 × Waist circumference (WC) + 0.874 × Body surface area (BSA) + 0.007 × Diastolic blood pressure (DBP) + 0.081 × Triglycerides (TG) - 0.445 × High-density lipoprotein cholesterol (HDL) + 0.306 × Fasting plasma glucose (FPG);Upper limit =  - 3.549 + 0.031 × Age + 0.339 × Sex + 0.030 × Waist + 2.721 × BSA + 0.020 × DBP + 0.004 × TG - 1.244 × HDL + 0.447 × FPG.The segmentation model achieved a Dice Similarity Coefficient of 0.926 and an ICC of 0.954(95 % CI: 0.946-0.961).In validation sets for diabetes and hypertension (n = 3,266), the screening detected 62.31 % abnormal cases, while in adrenal hyperplasia (n = 240), the rate was 77.09 %. CONCLUSION: This study developed individualized adrenal gland volume reference ranges and a low-dose CT-based deep learning tool for automated measurement. The screening tool shows potential to assist in identifying adrenal abnormalities and may provide a methodological basis for future clinical evaluation and application.