An AI-Based Algorithm for the Automatic Classification of Thoracic Radiographs in Cats.

Area of Science:

• Veterinary Radiology
• Artificial Intelligence in Medicine
• Machine Learning for Medical Imaging

Background:

• Radiographic interpretation of feline thoracic diseases is crucial for diagnosis.
• Developing automated tools can aid veterinary professionals in identifying common radiographic findings.
• Previous studies have explored AI for canine thoracic imaging, but feline applications require specific validation.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI)-based computer-aided detection (CAD) algorithm for common radiographic findings in the feline thorax.
• To compare the performance of two deep learning architectures, ResNet 50 and Inception V3, in detecting these findings.
• To assess the diagnostic accuracy of the AI algorithm across various thoracic conditions.

Main Methods:

• A multi-label convolutional neural network (CNN) approach was employed to build the CAD algorithm.
• The training database comprised feline thoracic radiographs from two institutions, including findings such as no findings, bronchial pattern, pleural effusion, mass, alveolar pattern, pneumothorax, and cardiomegaly.
• The performance of ResNet 50 and Inception V3 architectures was compared using the area under the receiver operating characteristic curve (AUC).

Main Results:

• Both ResNet 50 and Inception V3 architectures achieved high AUC values (above 0.9) for detecting alveolar pattern, bronchial pattern, and pleural effusion.
• AUCs above 0.8 were recorded for 'no findings' and pneumothorax.
• The algorithm demonstrated moderate performance for cardiomegaly (AUC > 0.7) and low performance for mass detection (AUC > 0.5) for both architectures. No significant differences in diagnostic accuracy were observed between the two CNN architectures.

Conclusions:

• The developed AI-based CAD algorithm shows promising diagnostic accuracy for several common feline thoracic radiographic findings, particularly pleural effusion, bronchial, and alveolar patterns.
• The algorithm's performance in detecting masses requires further improvement.
• Both ResNet 50 and Inception V3 architectures exhibit comparable performance, suggesting their suitability for this application in veterinary radiology.