AI in Veterinary Diagnostics: Analysing Physical Symptoms for Canine Health Assessment
Mbishibishi, Mugiraneza Leon Pierre (2025)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025052716690
https://urn.fi/URN:NBN:fi:amk-2025052716690
Tiivistelmä
Early detection of dermatological diseases in dogs is critical for timely treatment, yet traditional manual diagnosis is time consuming and resource intensive. Advancements in artificial intelligence (AI) have significantly impacted various domains, including veterinary medicine. The ability to analyse visual data using computer vision has opened new avenues for diagnosing and monitoring animal health conditions. Among these, dermatological conditions. This thesis explores an automated, image-based diagnostic pipeline leveraging computer vision and Vision Transformers (ViTs) to classify four common canine skin conditions (fungal infections, bacterial dermatosis, hypersensitivity dermatitis, and healthy skin).
A dataset of 1,831 labeled images was preprocessed and augmented (brightness ± 20 %, exposure ± 10 %) to increase model robustness and expand training data. Transfer learning techniques were applied, and the model was evaluated using overall accuracy and confusion matrices. The best model achieved 94.8 % test accuracy, with precision and recall above 90 % across all classes.
To ensure practical usability, the model was deployed as a cloud-hosted API on Roboflow platform, enabling instant inference prediction. This work demonstrates that ViT-based systems can deliver reliable, scalable dermatological diagnostics in resource-limited veterinary settings and lays the foundation for integration into telemedicine platforms and clinical decision support tools.
A dataset of 1,831 labeled images was preprocessed and augmented (brightness ± 20 %, exposure ± 10 %) to increase model robustness and expand training data. Transfer learning techniques were applied, and the model was evaluated using overall accuracy and confusion matrices. The best model achieved 94.8 % test accuracy, with precision and recall above 90 % across all classes.
To ensure practical usability, the model was deployed as a cloud-hosted API on Roboflow platform, enabling instant inference prediction. This work demonstrates that ViT-based systems can deliver reliable, scalable dermatological diagnostics in resource-limited veterinary settings and lays the foundation for integration into telemedicine platforms and clinical decision support tools.