Abstract:

Canine ocular diseases present significant diagnostic challenges in resource-limited veterinary settings due to overlapping symptoms and the scarcity of specialized equipment. This study addresses the urgent need for automated, accessible triage tools by benchmarking two distinct deep learning architectures: the Convolutional Neural Network (CNN)-based YOLOv8 and the Transformer-based RT-DETR. Utilizing a dataset of 1,557 clinical images curated via a novel "Shrink-Wrap" annotation protocol to mitigate background noise, we evaluated detection performance for Cataracts, Cherry Eye, and Conjunctivitis. The experimental results reveal a critical trade-off between computational efficiency and diagnostic sensitivity. While the lightweight YOLOv8n achieved rapid inference speeds (1.49 hours training time) suitable for mobile edge deployment, the RT-DETR model demonstrated superior efficacy in detecting diffuse inflammatory conditions. Specifically, the Transformer architecture achieved a 3.1% performance gain in Conjunctivitis detection (mAP@50 0.912) and significantly reduced background false positives from 10% to 3%. Furthermore, we implemented a heterogeneous ensemble using Weighted Boxes Fusion (WBF) to serve as a clinical safety net, minimizing false negatives. This research proposes a dual-deployment framework—utilizing YOLOv8 for offline mobile screening and RT-DETR for cloud-based confirmation—offering a scalable, cost-effective solution to bridge the diagnostic gap in rural veterinary practice in the Philippines.