Automated AI-based segmentation of canine hepatic focal lesions from CT studies.

Abstract

INTRODUCTION: Hepatic masses are a common occurrence in veterinary medicine, with treatment options largely dependent on the nature and location of the mass. The gold standard treatment involves surgical removal of the mass, often followed by chemotherapy if necessary. However, in cases where mass removal is not feasible, chemotherapy becomes the primary treatment option. Accurate lesion segmentation is crucial in such scenarios to ensure precise treatment planning. METHODS: This study aimed to develop and evaluate a deep learning-based algorithm for the automatic segmentation of hepatic masses in dogs. To achieve this, 200 canine CT cases with hepatic masses were collected from two clinics and the Antech Imaging Solutions database. Experienced veterinarians manually segmented the lesions to provide ground truth data. 25/200 CTs were excluded because they did not met the inclusion criteria. Finally, the algorithm was built using the nnUNet v2 framework and trained on 130 cases with a 5-fold training scheme. It was subsequently tested on 45 cases. RESULTS: The algorithm demonstrated high accuracy, achieving an average Dice score of 0.86 and an Average Symmetric Surface Distance (ASSD) of 2.70 mm. CONCLUSIONS: This represents the first report of a deep learning-based algorithm for the automatic segmentation of hepatic masses in dogs using CT imaging, highlighting its potential utility in clinical practice for improved treatment planning.