CSF-net: a color space fusion network with self-attention-driven feature learning for feline ocular diseases classification.

Introduction: Feline ocular diseases can cause irreversible vision loss if they are not detected early. However, early diagnosis is often difficult. This is due to limited access to veterinary ophthalmology services and the challenge of distinguishing between visually similar eye conditions. Illumination changes, glare, and strong visual similarity among diseases substantially limit the performance of conventional RGB-based classification models. This paper proposes the Color Space Fusion Network (CSF-Net), an attention-guided color interaction framework for robust feline ocular disease classification in real-world environments. Methods: CSF-Net decomposes images into complementary color spaces (RGB, HSV, and YCbCr), learns independent feature representations, and applies a self-attention mechanism to emphasize disease-relevant visual cues while reducing sensitivity to environmental variations. The proposed method is evaluated on a large-scale dataset of 16,480 feline ocular images and compared with representative deep learning models, including ResNet50, EfficientNet, and ViT/16-b. Results: CSF-Net achieves superior performance, reaching an accuracy of 83.05% and achieved a mean fold-wise macro-AUC of 0.9690 across five cross-validation folds. Ablation and class-wise analyses confirm the effectiveness of self-attention in leveraging multi-color representations, while also revealing limitations for eyelid-centered conditions such as blepharitis. A mobile-based preliminary screening application, PurrfectEyes, is further presented to demonstrate practical applicability. Discussion: Overall, this work introduces an attention-guided multi-color framework that improves robustness and interpretability for AI-assisted feline ocular disease screening in real-world settings.