Vetnews | Oktober 2025 22 « BACK TO CONTENTS These comparisons highlight the strengths and weaknesses of different approaches, allowing for a comprehensive assessment of model efficacy in real-world applications. 4.1 Test dataset The test dataset for evaluating the U-Net, GPT-4, and Grounding SAM models was meticulously assembled to ensure diversity and clinical relevance, drawing from a wide range of veterinary resources. The images in the dataset were sourced from veterinary medical encyclopedias, peer-reviewed medical articles, and specialised veterinary clinics, all of which provided documented diagnoses and necessary permissions [53]. Each image in the dataset was resized to a standardised 320x320 pixel format to facilitate consistent processing and analysis across models. This standardisation is crucial to maintain the uniformity of input data and ensure the comparability of model outputs during testing. A representative example from the dataset is shown in Figure 19. includes images of an American bulldog [97] and a crossbred dog [98], both diagnosed with Cherry Eye as seen in. These images depict a pink-colored ocular protrusion from the medial canthus—a classic sign of Cherry Eye—accompanied by excessive tearing. Although both dogs present similar symptoms, their differing breeds introduce subtle anatomical variations that challenge automatic segmentation models. The models must learn to handle such variations and correctly interpret symptoms in diverse breeds. This illustrates some of the inherent difficulties that computer vision must address when attempting to accurately segment and diagnose eye conditions in canines. The test dataset was annotated in detail to reflect four key ocular symptoms that are commonly observed in veterinary ophthalmology: Corneal Cloudiness or Haziness, Sclera Redness, Excessive Tearing and Colored Mass Protrusion in the Corner of the Eye. Each image was paired with a one-channel mask, stored in PNG format, annotating the four distinct classes of symptoms. These annotations are crucial for training and validating the models, as they provide ground truth data that allows the models to learn the correct segmentation patterns. The pixel intensity in the mask varies depending on the symptom being represented, allowing for precise identification of each condition. This approach mirrors the methodology used for the DogEyeSeg4 custom dataset, ensuring consistency between training and test datasets. The dataset includes images representing five significant ocular diseases in canines: Cherry Eye, Glaucoma, Uveitis, Corneal Ulceration, and Bacterial Keratitis. Each of these diseases is featured through five distinct cases, providing a total of 25 images for comprehensive model testing. • Cherry Eye: This condition is characterised by the prolapse of the gland of the third eyelid, visible as a pink-colored mass in the corner of the eye. Cherry Eye can cause discomfort and, if left untreated, may lead to more severe complications such as dry eye or infection. The test images show various stages of Cherry Eye, providing the models with different levels of protrusion and tear production to segment and interpret. • Glaucoma: Caused by elevated intraocular pressure, Glaucoma often results in pain, corneal cloudiness, and scleral redness. The dataset includes images showing a range of Glaucoma severity, from mild to advanced cases, where the eyeball is visibly enlarged. These images challenge the models to recognise subtle differences in redness intensity and cloudiness levels. • Uveitis: This inflammatory condition affects the uveal tract and often presents with scleral redness, pain, and sensitivity to light. In some cases, corneal cloudiness is also observed. The Uveitis cases in the dataset emphasise the need for precise segmentation of redness and cloudiness, ensuring that the models can distinguish between inflammatory conditions and other ocular issues. • Corneal Ulceration: This condition is typically the result of trauma or infection and is marked by corneal erosion or scratches. The test images highlight the symptoms of excessive tearing and sclera redness, along with visible cloudiness in some cases. These images require the models to focus on small, localised areas of damage, a challenge in real-world clinical scenarios where the ulcers may be difficult to detect without specialised equipment. • Bacterial Keratitis: Bacterial infection of the cornea leads to symptoms such as corneal cloudiness and tearing and can progress rapidly if untreated. The dataset provides examples of varying severity, allowing. The models to learn how to segment the infected areas accurately and distinguish them from less severe conditions like Corneal Ulceration. Each disease class is represented with diverse images, ensuring that the models are exposed to a broad spectrum of cases. This enables the models to generalise better and handle real-world variations in symptom presentation, breed anatomy, and disease progression. These images, along with their corresponding annotated masks, form the foundation for the final model evaluation. The sample of the test dataset can be observed in Figure 20. [63]. This test dataset is not only a tool for evaluating the U-Net segmentation model but also serves as a benchmark for comparing the performance of the GPT-4 and Grounding SAM models. By utilising the same dataset across all models, the results can be compared directly, allowing for a thorough performance evaluation. The models will be assessed based on their ability to accurately segment and interpret the symptoms associated with each disease, with particular focus on their precision in identifying small, nuanced differences between similar conditions. The goal of this evaluation is to determine which model offers the most effective solution for automated segmentation and diagnosis in canine ophthalmology. The evaluation will be detailed further in the “Performance Evaluation” section, where we will outline the specific metrics and outcomes that support this comparative analysis. Leading Article Figure 19. Images of an American bulldog (left) and a crossbred dog (right) show a pink-colored mass protruding from the medial canthus of both eyes, accompanied by signs of inflammation, including pain and excessive tearing, indicative of Cherry Eye
RkJQdWJsaXNoZXIy OTc5MDU=