Deep Learning Model for CT-based Adrenal Gland Volume Determination and Normal Reference Definition in Dogs

Adrenal gland size is linked to its function, disease status, and tumor malignancy, if any, making accurate measurement of its size essential. However, measuring adrenal gland length is prone to errors, and volume is a reliable indicator of its size. Manual volume measurement is time-consuming and is usually inaccurate. Therefore, this study aimed to develop an artificial intelligence (AI) model for direct adrenal gland volume measurement in computed tomography (CT) images. Post-contrast CT images of 250 dogs were segmented. Of these, 200 scans were randomly selected for training and 50 for validation. A deep learning model, based on Swin-Transformers and several processing techniques, was developed. Computed tomography images of 239 dogs were used for normal reference definition, with adrenal gland volume was determined on the basis of the absence of adrenal gland lesions supported by clinical and laboratory data. The mean (±SD) Dice Similarity Coefficient (DSC) of adrenal gland segmentation was 0.885±0.075, which is slightly lower than other abdominal organs of dogs, most probably due to the small size, varied shapes, and overlapping with surrounding tissue. Agreement analysis between manual voxel counts and the AI model showed an interclass correlation coefficient of 0.957 (P<0.001). Adrenal gland volume correlated positively with body weight (BW; r=0.821, P<0.001) and age (r=0.147, P<0.05), and negatively with body condition score (BCS; r=-0.233, P<0.001). The relationship was represented by the regression equation: adrenal volume=-0.51xBCS+0.033×BW+0.015×age+0.373 (=0.72, P<0.001). No correlation was found between adrenal gland volume and sex of dogs. In conclusion, an AI model was developed to directly measure adrenal gland volume from CT images of dogs, which would potentially aid in adrenal disease screening.