Automated segmentation of canine pulmonary masses in CT imaging using AI

Primary pulmonary lung cancer is rare in dogs, and clinicians increasingly rely on advanced imaging for diagnosis and treatment planning. However, manual lesion segmentation can be time-consuming and subject to operator variability. This retrospective study compiled a multicenter dataset of canine CT scans containing at least one pulmonary mass measuring more than 2 cm. Data were collected from two university veterinary hospitals and a teleradiology service, encompassing varying acquisition protocols and scanner types. Lesions were manually segmented to create ground truth masks, and an AI model was trained and evaluated using the nnUNet v2 framework with a 5-fold cross-validation approach. Performance on a separate test set of 30 scans was quantified using the Dice Similarity Coefficient (DSC) and Average Symmetric Surface Distance (ASSD). The databse was made of 217 cases. The training/validation set comprised 187 cases. The model's segmentation accuracy was tested on 30 cases. The trained model had a high segmentation accuracy on the test set, with a mean DSC of 0.91 and an ASSD of 1.88 mm. The model had high performance on homogeneous, well-defined masses, whereas the presence of intralesional mineralisation or pleural effusion had a negative impact on the model's performance.