VN October 2025

Vetnuus | October 2025 21 within the canine images, confirming its effectiveness for this specific application in veterinary ophthalmology. While the model demonstrates strong performance metrics, occasional detection inaccuracies underscore the need for ongoing adjustments and improvements in model training. 3.3. Web Application Development The development of the web application for eye symptoms segmentation and medical condition prediction focuses on providing a robust, user-friendly interface to assist veterinarians and pet owners in diagnosing canine ocular diseases using AIbased image analysis. The web application is built to integrate a suitable segmentation model, ensuring seamless deployment, accessibility, and scalability. The development process is initiated by designing an intuitive user interface that would guide users through the image acquisition and diagnostic process. As seen in Figure 17, the system starts with a user interface that allows users to either capture images of their dog’s eyes or upload them from local storage. This flexibility accommodates both veterinary professionals and pet owners who may not have prior experience with such diagnostic tools. The quality and framing of these images are crucial for accurate segmentation [56], so the interface should include an overlay mask as shown in Figure 18.[54], guiding users on how to correctly align the dog’s eye for optimal image capture [54]. The backend architecture of the application was designed to support the high computational demands of the segmentation model while ensuring portability. The system is containerised using Docker [91], which simplifies the deployment process and ensures that the application runs consistently across different platforms. This containerised approach also allows the application to scale efficiently, accommodating an increasing number of users without sacrificing performance. The core image processing tasks are handled by a suite of Python [92] libraries, including TensorFlow [93] for model execution, OpenCV [94] for image manipulation, and Pillow [95] for image formatting. The backend server, built using Flask [96], ensures stable communication between the web interface and the desired segmentation model, processing images and delivering results in real-time [54]. The entire setup ensures that the application is portable, scalable, and easy to deploy on both local machines and cloud-based servers. To ensure efficient handling of requests and model processing, the deployment was optimised for server environments equipped with advanced graphics processing units (GPUs), ensuring that the heavy computational tasks, such as segmenting and analysing canine eye images, are handled with minimal latency. This serverbased approach also allows for future scalability, where additional images can be collected and processed, contributing to the improvement of the model through continuous retraining and dataset expansion. Additionally, the application is built with the capacity to expand into mobile platforms, further enhancing its accessibility. The web application not only provides diagnostic insights but also aims to enrich the model by collecting user data with their consent. Each user interaction and image submission can potentially contribute to the growth of the dataset, allowing for further model refinement. This continuous loop of data collection, analysis, and improvement represents the application’s potential for evolving alongside the veterinary community’s needs. The user experience is further enhanced with feedback features, guiding users with next steps based on the results, and providing valuable insights into their pet’s ocular health. The combination of these technical and practical design choices has led to a comprehensive tool for diagnosing canine eye diseases, setting the foundation for further developments and improvements. 4. EXPERIMENTS AND RESULTS This section presents a detailed analysis of the experimental procedures and outcomes, focusing on the performance of various models in diagnosing canine ocular diseases. It is structured to provide insights into the evaluation process through several stages. First, we describe the test dataset used for model evaluation, including its structure and content. Next, we elaborate on the metrics applied to assess both segmentation and text generation tasks. The core of the experimentation focuses on the training and evaluation of the U-Net model, both with and without backbone networks. A comparative study is also performed between the U-Net model trained on a real dataset versus one trained on a combined real and synthetic dataset to evaluate the impact of synthetic data augmentation. In a separate experiment, the benefits of binary vs. multiclass models were assessed along with their limitations. Additionally, this section compares the performance of U-Net against GPT-4 and Grounding Dino in both segmentation and diagnostic interpretation tasks. Leading Article Figure 17. A web-based application flowchart for eye symptoms segmentation and medical condition prediction Figure 18. An overlay mask acting as a capturing images guide for accurate segmentation >>>22

RkJQdWJsaXNoZXIy OTc5MDU=