Deep learning-based diagnosis of feline hypertrophic cardiomyopathy

Area of Science:

• Veterinary Medicine
• Artificial Intelligence
• Cardiology

Background:

• Feline hypertrophic cardiomyopathy (HCM) affects 10-15% of cats, causing serious health issues.
• Radiography and ultrasound are diagnostic gold standards, but radiography alone has limited accuracy (75%).
• Advanced diagnostic tools are needed to improve feline HCM screening.

Purpose of the Study:

• To investigate the optimal deep learning architecture for diagnosing feline HCM via radiography.
• To evaluate the accuracy of various residual neural network architectures in detecting feline HCM.
• To assess the effectiveness of voting strategies in improving diagnostic accuracy.

Main Methods:

• Trained five residual architectures (ResNet50V2, ResNet152, InceptionResNetV2, MobileNetV2, Xception) on 231 ventrodorsal radiographic images (143 HCM, 88 normal).
• Utilized a diverse dataset from 5 independent institutions to ensure generalizability.
• Applied softmax and majority voting strategies to combined model predictions.

Main Results:

• All five models achieved over 90% accuracy, with ResNet50V2, ResNet152, InceptionResNetV2, MobileNetV2, and Xception all reaching 95.45%.
• Softmax voting strategy achieved 95% accuracy on combined test data.
• The deep learning system demonstrated high potential for assisting in feline HCM diagnosis.

Conclusions:

• Automated deep learning systems using residual architectures show high accuracy in diagnosing feline HCM from radiographs.
• These AI tools can significantly assist veterinary radiologists in screening for feline hypertrophic cardiomyopathy.
• Further development of AI in veterinary diagnostics can improve early detection and management of feline heart disease.