Classification of Left Atrial Enlargement in Canine Radiographs Using Artificial Intelligence

Myxomatous mitral valve disease (MMVD) accounts for about 75% of heart disease cases in dogs. It involves mitral valve degeneration, often progressing to mitral regurgitation and congestive heart failure. One of the main elements for its diagnosis is chest radiography, which allows the assessment of key symptoms, such as an enlarged left atrium. However, it requires specialized expertise that many veterinary clinics lack. Combined with the high volume of X-rays, this can frequently lead to misdiagnoses. In this study, we propose an AI-based model to help classify an enlarged left atrium in canine chest X-rays, aiding MMVD diagnosis. We created a dataset of lateral chest X-rays, comprising normal patients (n=290) and those with an enlarged left atrium (n=160) and trained several Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) architectures. Data augmentation was applied to improve model robustness, and a 5-fold cross-validation assessed performance across accuracy, precision, sensitivity, specificity, F-score, and AUC metrics. We combined the best models obtained to create ensembles. The best ensemble, consisting of InceptionV3, VGG19, and DenseNet-121 architectures, achieved an F-score of 0.8892 ± 0.0371 and an AUC of 0.9099 ± 0.0508, offering a balanced classifier that enhanced true positive detection while reducing false positives.