VN October 2025

Vetnews | Oktober 2025 12 « BACK TO CONTENTS veterinary care and no animals were harmed for the purpose of image acquisition, ethical approval was not required for this study. This was made explicit in the ethics and consent statement accompanying the dataset: “Images were collected as part of routine clinical evaluations, and no ethical approval was necessary as no harm was caused to the animals.” One of the key challenges in assembling this dataset was balancing the representation of each class. For example, some symptoms, such as excessive tearing (S3), appeared more frequently than others, such as the colored mass (S4). Without careful selection, this imbalance could lead to biased model training, where the model becomes overly tuned to detecting the more frequent symptoms. Therefore, a meticulous curation process was employed to ensure a balanced distribution across all four classes. To further enhance the dataset’s utility, data augmentation techniques were applied. Building on methodologies from medical image segmentation, particularly with respect to U-Net architecture, augmentation methods such as horizontal flipping, rotation by up to 15 degrees, and translation by 50 pixels were used as seen in Figure 3. [56]. These augmentations introduce variation in camera angles and positions, simulating different real-world scenarios where veterinary professionals might capture images from slightly different perspectives. Importantly, zoom augmentation was avoided to prevent interpolation, which could introduce noise into the masks and degrade the accuracy of the annotations. The augmentation process expanded the dataset to 200 images, increasing its robustness and making it more suitable for training machine learning models that need to generalise across diverse imaging conditions. 3.1.2 Synthetic datasets Due to the limited availability of real-world data and the need for large datasets to train deep learning models, synthetic image generation has become a valuable resource in medical imaging, including veterinary ophthalmology. Several techniques exist for generating synthetic images, each offering strengths and weaknesses in tackling this issue. These methods include Generative Adversarial Networks (GANs) [57], Variational Autoencoders (VAEs) [58], and Diffusion Models [59]. Additionally, procedural methods like rule-based systems and 3D rendering techniques are also applied for synthetic image generation, offering high control over image features, albeit with increased manual intervention and resource demand [60]. GANs utilise a two-part system, where the generator creates images, and the discriminator attempts to distinguish them from real images. While GANs are capable of producing high-resolution, realistic images, they are computationally expensive and often unstable during training, with issues like “mode collapse,” where limited variations of images are generated. VAEs, by contrast, are probabilistic models that encode input data into a latent space and then decode it back, generating new data by sampling from this latent space. VAEs are more stable to train than GANs but tend to produce lower-resolution images that may lack the detailed features required in tasks such as ophthalmic disease diagnosis. Diffusion models, particularly the Stable Diffusion variant used in this study, offer a more balanced approach, combining computational efficiency with high-quality image generation. A significant feature of diffusion models is their reliance on the U-Net architecture. In diffusion models, the U-Net serves as the backbone for the denoising process, transforming random noise into coherent images. This structure makes diffusion models particularly well-suited for generating structured medical images [61]. Beyond learning-based techniques, rule-based image generation and 3D rendering techniques are also used in certain fields for procedural image creation [60]. Rule-based systems apply predefined algorithms to generate images, offering high customisation but requiring extensive manual intervention. Rendering techniques simulate 3D environments and lighting to create highly detailed images, but these methods demand considerable setup and computational resources [62]. Figure 3. Example of applied augmentations on original image in top row: horizontal flip (middle row left), horizontal shift (middle row right), rotation (bottom row left), and vertical shift (bottom row right) Figure 4. Stable diffusion generated images with fixed posture of different dog breeds. Images shows high degree of reality with certain difficulty like double set of ears Leading Article

RkJQdWJsaXNoZXIy OTc5MDU=