Artificial intelligence evaluating primary thoracic lesions has an overall lower error rate compared to veterinarians or veterinarians in conjunction with the artificial intelligence

Abstract

To date, deep learning technologies have provided powerful decision support systems to radiologists in human medicine. The aims of this retrospective, exploratory study were to develop and describe an artificial intelligence able to screen thoracic radiographs for primary thoracic lesions in feline and canine patients. Three deep learning networks using three different pretraining strategies to predict 15 types of primary thoracic lesions were created (including tracheal collapse, left atrial enlargement, alveolar pattern, pneumothorax, and pulmonary mass). Upon completion of pretraining, the algorithms were provided with over 22000 thoracic veterinary radiographs for specific training. All radiographs had a report created by a board-certified veterinary radiologist used as the gold standard. The performances of all three networks were compared to one another. An additional 120 radiographs were then evaluated by three types of observers: the best performing network, veterinarians, and veterinarians aided by the network. The error rates for each of the observers was calculated as an overall and for the 15 labels and were compared using a McNemar's test. The overall error rate of the network was significantly better than the overall error rate of the veterinarians or the veterinarians aided by the network (10.7%vs 16.8%vs17.2%, P=.001). The network's error rate was significantly better to detect cardiac enlargement and for bronchial pattern. The current network only provides help in detecting various lesion types and does not provide a diagnosis. Based on its overall very good performance, this could be used as an aid to general practitioners while waiting for the radiologist's report.