Diagnosing Brachycephalic Obstructive Airway Syndrome in Dogs Using Computer Vision and Machine Learning

Abstract

Brachycephalic Obstructive Airway Syndrome (BOAS) is a breathing disorder that is common among dogs of certain breeds, impairing their quality of life. Diagnosing BOAS requires licensed veterinarians to undergo specialized training, which limits accessibility. There is demand for a more accessible solution, which machine learning could provide. This thesis evaluates two different approaches of machine learning to explore the possibility of classifying BOAS based on audio recorded from Android mobile phones, using Python. The first approach extracts signal features from the audio, which are used to train a simple machine learning model. The second approach relies on computer vision, using spectrograms; visual representations of the audio signals, to train a Convolutional Neural Network (CNN). Due to limited data, I employed augmentation techniques to artificially expand the dataset for training the models. My findings suggest that the spectrogram-based model is better suited for the problem, with a perfect prediction accuracy when the dogs were recorded after a short exercise test, suggesting strong performance under that condition. For dogs at rest, this model achieved an 81.7% accuracy, indicating somewhat promising results even under less favourable conditions. However, due to the limited dataset, the predictive performance was evaluated on few samples, and therefore additional data is needed for a more robust conclusion. Furthermore, one particular augmentation technique designed to account for differences in the recording devices’ frequency response enhanced the model’s general performance and could be further refined to improve accuracy, especially for dogs at rest.