This study explores the integration of a specialized U-Net model and large language models (LLMs) for diagnosing canine eye diseases. The U-Net model, enhanced with various backbones such as ResNet34, Inception V3, VGG16, and EfficientNet B3, was trained on a custom dataset curated by veterinary ophthalmology specialists. The model effectively segments images based on key symptoms: Ocular Opacity, Sclera Redness, Excessive Tearing, and Colored Ocular Protrusion. These segmented images are interpreted using LLMs, including ChatGPT, Mistral, Gemini, Llama, and Claude, to provide preliminary diagnoses. Evaluation metrics, such as MPNet, MiniLM, BERTScore, CLIPScore, BLEU, METEOR, ROUGE, and SPICE, were employed to assess the accuracy of the LLMs in generating diagnostic insights from the segmented images. The study found that the U-Net model with ResNet34 backbone achieved the highest accuracy in identifying sclera redness, with a Jaccard index of 80.6%. Among the LLMs, GPT4 exhibited the best performance across all metrics in both single and dual diagnosis scenarios. The integration of U-Net and LLMs demonstrated improvements in diagnostic accuracy and efficiency, suggesting a viable approach for veterinary diagnostics.