Vetnews | Oktober 2025 10 « BACK TO CONTENTS Leading Article providing high accuracy in ocular image analysis and disease classification [1], [2]. These advancements are attributed to CNNs’ ability to automatically extract and learn relevant features from ocular images, improving diagnostic accuracy compared to traditional manual methods [3], [4], [5]. The development of these CNN-based models has been supported by a variety of publicly available datasets, which have enabled the training and validation of models in human ophthalmology. Notable datasets include: ORIGA-light dataset focusing on optic nerve head segmentation and glaucoma assessment [6], DrishtiGS, designed for glaucoma detection, containing annotated fundus images for training CNNs [7], Retinal Fundus Image for Glaucoma Detection with a focus on glaucoma diagnosis [8], RIMONE, a large dataset for retinal image analysis, primarily used for glaucoma detection [9], iBUG focuses on facial landmark detection but has also been used for eye-tracking and ocular disease analysis [10], OpenEDS dataset that includes annotated eye images for eye-tracking and ocular disease applications [11], UBIRIS primarily for biometric purposes, but also applied in ocular disease diagnosis [12] and TEyeD dataset containing eye-tracking data, useful for studying eye movements and diagnosing conditions such as glaucoma [13]. These datasets have significantly contributed to the development of effective models for human ophthalmology. In contrast, the application of computer vision techniques in veterinary ophthalmology, particularly for canine eye diseases, is much less developed. Datasets and research in this field are scarce. Studies focusing on conditions such as canine glaucoma are limited [14], [15], [16], and the few available datasets are typically small and lack the diversity seen in human datasets. A notable study used CNNs to diagnose ulcerative keratitis in dogs, but it was constrained by a limited dataset and acquired under controlled conditions [17]. This highlights the need for more extensive datasets and research into canine ophthalmology. The U-Net architecture, introduced for biomedical image segmentation, has become a popular choice for various medical imaging tasks due to its encoder-decoder structure, which enables precise feature extraction and localisation [18]. U-Net’s architecture is particularly effective in segmenting medical images where pixel-level accuracy is critical. It has demonstrated robust performance across multiple domains, including ophthalmology, where it has been applied to tasks such as optic disc segmentation, retinal layer segmentation, and detecting diseases like glaucoma and cataracts [19], [20], [21], [22]. Studies have consistently shown that U-Net performs well in medical image analysis, confirming its relevance in both human and veterinary medical contexts [18], [19], [20], [23], [24], [25]. Transformer-based models, such as Swin Transformer [26] and SegFormer [27], have recently gained attention for their superior ability to capture global context, making them particularly effective in fine-grained segmentation tasks across human and veterinary fields. However, U-Net continues to be a formidable option, especially in situations with limited data availability, where it still performs robustly [28]. In fact, a comparison between U-Net and transformer-based architectures for medical image registration demonstrated that UNet, with minimal adjustments, can surpass the performance of these newer transformer models [28]. One of the key strengths of U-Net is its adaptability, particularly when combined with transfer learning. Transfer learning allows models to leverage pre-trained weights from large datasets, enabling them to perform well even on smaller datasets typical in veterinary applications. U-Net has been successfully combined with advanced CNN backbones like ResNet [29], EfficientNet [30], VGG [31], and Inception [32], significantly improving its feature extraction capabilities [4], [5]. These backbones help enhance the model’s ability to learn from limited data, making it well-suited for applications where obtaining large datasets is impractical. Furthermore, studies have shown that U-Net does not require an extensive dataset to achieve good results, especially when augmented with transfer learning techniques [33]. While U-Net can perform well with small datasets, the creation of synthetic datasets can further enhance model performance by augmenting the available training data. Synthetic data generation techniques, such as those using Generative Adversarial Networks (GANs) and diffusion models, have become valuable tools for augmenting datasets in various domains [34], [35]. Diffusion models, particularly Stable Diffusion, have shown promise in generating high-quality synthetic images by iteratively refining noisy images [36]. Stable Diffusion, which employs a U-Net-like architecture for image synthesis, can be used to create realistic synthetic images of diseased eyes, which can significantly improve the training of CNNs for canine ophthalmology [37], [38], [39]. These synthetic datasets could address the shortage of annotated data in veterinary ophthalmology and enhance model robustness [40]. A web-based application utilising a U-Net model trained on real and synthetic datasets could provide a valuable diagnostic tool for veterinarians. This application could assist in the early detection and treatment of canine eye diseases. In combination with image segmentation, the integration of Large Language Models (LLMs) could further enhance the diagnostic process. LLMs such as ChatGPT [41], Mistral [42], Gemini [43], Llama [44], and Claude [45] have demonstrated significant potential in medical data interpretation. These models, when integrated with image analysis tools, could help in interpreting symptoms, guiding veterinarians through diagnostic workflows, and improving decision-making. Evaluation metrics like BERTScore [46], CLIPScore [47], BLEU [48], METEOR [49], ROUGE [50], and SPICE [51] have shown that LLMs can effectively process complex medical information, making them suitable for integration into diagnostic applications. This research focuses on developing a mobile web application for diagnosing canine eye diseases by combining U-Net-based image segmentation with LLMs for interpreting medical symptoms. The U-Net model is trained on a custom dataset – DogEyeSeg4 [52], augmented with synthetic data generated using Stable Diffusion techniques. This study explores the integration of LLMs with medical image analysis, evaluating the potential of various LLMs such as ChatGPT, Mistral, Gemini, and others in improving diagnostic workflows. By combining these state-ofthe-art technologies, this research aims to advance the field of veterinary ophthalmology, providing tools that can assist in the early detection and treatment of eye diseases in dogs. 3. METHODOLOGY 3.1. Dataset In veterinary ophthalmology, especially in the niche of canine eye diseases, obtaining suitable datasets poses significant challenges.
RkJQdWJsaXNoZXIy OTc5MDU=