Vetnews | Oktober 2025 26 « BACK TO CONTENTS The published custom-made DogEyeSeg4 dataset was important in training the models. It provided a diverse set of images reflecting real-world clinical conditions, including variations in lighting, breed, and image quality. The inclusion of synthetically generated images expanded the dataset, allowing models to learn from a broader range of symptom presentations. This comprehensive dataset contributed significantly to the robustness and generalizability of the models. The limitations identified are the model’s sensitivity to image quality, where the diagnostic accuracy is highly dependent on the quality of input images and segmentation masks. Poor image quality or inaccurate segmentation can negatively impact model performance. Also noticed is that models occasionally struggled with conditions like Corneal Ulcer, suggesting the need for further refinement in handling less visually distinct symptoms. Some LLMs, like Llama-3, underperformed in diagnostic interpretation, indicating variability in model capabilities. Future research should focus on extending the dataset by incorporating more diverse images and medical symptoms, including rare conditions and different breeds, to improve model generalisation. Further usage of advanced model architectures, such as transformerbased models for segmentation to capture global contextual information, could be beneficial. Multimodal data integration by combining image data with clinical history and other modalities could provide more info for a more holistic diagnostic approach. Testing the diagnostic pipeline in clinical settings to validate and refine the models based on practical use should provide valuable feedback. The experimental results confirm both hypotheses: Hypothesis 1 is supported by the successful development of a computer vision U-Net(RSD) model for segmentation, and implementation of GPT-4o for interpretation, that accurately recognises canine ocular conditions in images taken in unconstrained environments. Hypothesis 2 is confirmed by demonstrating that modifications to the U-Net network’s input and architecture by using a combination of real and synthetic images, a custom attention mechanism in the form of heatmaps, integrating advanced backbones and employing combined loss functions, lead to better segmentation of canine eye conditions. 6. CONCLUSION This research has confirmed both initial hypotheses: Hypothesis 1: By developing a computer vision U-Net(RSD) model for segmentation and implementing GPT-4/GPT-4o for interpretation, we demonstrated that certain canine ocular conditions can be accurately recognised in still images taken in unconstrained environments. Hypothesis 2: Modifying the U-Net network’s input and architecture by using a custom attention mechanism (heatmaps), integrating advanced backbones, and employing combined loss functions resulted in improved segmentation of canine eye conditions. Leading Article Figure 40. The depicted diagnostic pipeline for canine ophthalmic conditions demonstrates improved performance when both the original images and their corresponding segmented masks from U-Net(RSD) are utilised as inputs for GPT-4o, enhancing the accuracy of the medical diagnoses derived from visual symptoms
RkJQdWJsaXNoZXIy OTc5MDU=