UNet for automatic pneumothorax detection on canine and feline CTs

Abstract Pneumothorax is defined as the pathological presence of air or gas in the pleural space. This can be due to penetration of airthrough a pleuro-cutaneous, pleuro-pulmonary or pleuro-esophageal pathway. Pneumothorax is a potentially life-threateningcondition and its detection in routine clinical and emergency settings can be decisive for patient survival. Computed tomography(CT) plays a key role in determining its presence and extent. This retrospective study aimed to develop a deep learning-basedalgorithm for automatic segmentation of pneumothorax in dogs and cats. Data were collected from different facilities. Thepathological air accumulation was then manually segmented by experienced radiologists to create a ground truth. An nnU-Netframework was used to build the algorithm. One hundred cases were collected, and the model was trained on 80 cases andtested on the remaining 20. The model was then tested on 47 negative cases. Performances were evaluated using Dicesimilarity score (DSC), the Aggregated Dice Score Similarity Coefficient (DSCAgg), and the Average Symmetric SurfaceDistance (ASSD). The model reached a good detection capability with a DSC of 0.797, a DSCAgg of 0.9267 and an ASSDof 3.281. This study is the first reporting the development of a deep learning-based algorithm for automatic segmentation ofpneumothorax in dogs and cats on CT scans, suggesting the potential impact in clinical and emergency scenarios.