Automatic classification and grading of canine tracheal collapse on thoracic radiographs by using deep learning

Area of Science:

• Veterinary Radiology
• Artificial Intelligence in Medicine
• Canine Respiratory Diseases

Background:

• Tracheal collapse is a progressive airway disease where symptom severity correlates with the degree of collapse.
• Interpreting radiographs for tracheal collapse presents challenges for veterinarians due to diagnostic uncertainties.
• Automated screening tools are needed to improve disease detection in veterinary settings.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) model for screening canine tracheal collapse.
• To differentiate between normal tracheas and varying degrees of tracheal collapse using radiographs.
• To assess the diagnostic performance and reliability of the AI model.

Main Methods:

• An AI model based on You-Only-Look-Once (YOLO) algorithms (v3, v4, v4 tiny) was trained and tested.
• The model utilized archived lateral cervicothoracic radiographs of dogs.
• The YOLO v4 tiny-416 model was specifically evaluated for its screening capabilities.

Main Results:

• The YOLO v4 tiny-416 model achieved 98.30% sensitivity, 99.20% specificity, and 98.90% accuracy in classifying tracheal collapse.
• The model successfully differentiated normal tracheas from Grade 1-2 and Grade 3-4 tracheal collapse.
• High diagnostic accuracy was confirmed with an area under the precision-recall curve >0.8.
• Excellent agreement (κ=0.975) and consistency (ICC >0.90) were observed between the AI model and radiologists.

Conclusions:

• The developed deep learning model is a reliable tool for screening and grading canine tracheal collapse.
• This AI approach can assist veterinarians in diagnosing tracheal collapse using standard radiographs.
• The model offers a valuable method for improving the efficiency and accuracy of canine airway disease assessment.