VN October 2025

Vetnuus | October 2025 11 Leading Article Unlike human medical datasets, which benefit from more standardised and widely available sources, veterinary datasets are relatively scarce. The images needed for training models in this domain must capture a variety of conditions across multiple breeds, often under nonideal circumstances. As a result, a custom dataset was developed to support this research, addressing the scarcity of available data. The following sections describe the DogEyeSeg4 dataset, constructed to overcome these challenges, as well as a synthetic dataset generated to augment the available real-world data. 3.1.1 DogEyeSeg4 Custom dataset Given the limited availability of publicly accessible canine ophthalmic datasets, the development of the DogEyeSeg4 dataset became essential. The process of gathering suitable data from real-world clinical environments posed several difficulties. First, patient compliance during eye examinations often impacted image quality. Dogs, being non-cooperative subjects, frequently moved during assessments, resulting in blurry images or requiring multiple attempts to capture usable data. Second, the diversity of breeds presented additional variability in eye structure, fur colouration, and size, all of which affected the clarity and focus of the images. These factors, while reflective of real-world conditions, made it difficult to obtain consistent, high-quality images that were necessary for robust model training. The images in the DogEyeSeg4 dataset were collected from two specialised veterinary ophthalmology clinics and a veterinary eye disease atlas [53]. Clinical environments, unlike controlled laboratory settings, introduce uncontrollable variables such as inconsistent lighting, non-standardised camera equipment, and varying angles of capture. These factors result in natural but challenging conditions for machine learning models. The lighting in veterinary clinics often varies depending on the location of the examination, leading to different levels of contrast and exposure in the images. Additionally, images were captured without staged conditions, meaning the dataset includes a range of natural scenarios rather than perfectly lit or artificially enhanced images. This variety adds to the complexity but also enhances the dataset’s applicability to real-world diagnostic conditions. The DogEyeSeg4 dataset consists of 145 images, which include both close-up images of the canine eye and full headshots. The example images with corresponding masks are shown in Figure 1. This diversity is crucial because veterinary practitioners often capture images that show either the entire head of the animal or focus specifically on the affected eye, depending on the diagnostic requirement. Close-up images provide detailed views of specific conditions, such as corneal cloudiness or excessive tearing, while whole-head images are more common in clinical settings and may capture multiple symptoms, including redness of the sclera or colored masses in the corner of the eye, from a broader perspective. Each image was resized to 320x320 pixels for consistency during model training, and the dataset is annotated with one-channel masks in PNG format. The four annotated classes, described in Figure 2.[54], correspond to the following symptoms: • S1: Cloudiness or haziness of the cornea, • S2: Redness of the sclera, • S3: Excessive tearing, • S4: A colored mass in the corner of the eye. These symptoms correspond to several common diseases, such as Cherry Eye, Glaucoma, Uveitis, Corneal Ulceration, and Bacterial Keratitis. To ensure the clinical relevance and accuracy of the dataset, each image and its annotation were reviewed by a veterinary specialist. This review process was critical in maintaining the diagnostic. Ensuring the dataset complied with data protection regulations, particularly the General Data Protection Regulation (GDPR) [55], was a critical part of the dataset creation process. All images included in the dataset were anonymised to safeguard the privacy of clients and their animals. Anonymisation involved the removal of any identifiable information, such as examination dates, client names, and animal identifiers. By adhering to GDPR guidelines, the dataset was rendered compliant with strict privacy regulations. Additionally, since the images were gathered during standard Figure 1. Example of images in DogEyeSeg4 dataset with corresponding masks showing closeup of an eye (upper row) and whole head (lower row) Figure 2. Visual representation of medical symptoms in the top-left to bottom-right order: S1 - cloudiness or haziness of the cornea; S2 - sclera redness; S3 - excessive tears; and S4 - colored mass protrusion in the corner of the eye. Precision of the dataset, especially given the variability in patient behaviour and the nonstandardized clinical conditions under which the images were captured

RkJQdWJsaXNoZXIy OTc5MDU=