VN October 2025

Vetnews | Oktober 2025 14 « BACK TO CONTENTS • Controlled Variability: Synthetic generation allows researchers to precisely control the inclusion of specific symptoms, conditions, and their severity. This level of control ensures that the dataset remains balanced, mitigating issues related to class imbalance, and comprehensively covering the spectrum of disease presentations [69]. • Ethical Benefits: The use of synthetic images circumvents the need for invasive clinical procedures or additional veterinary visits. As a result, it provides an ethically sound method for expanding datasets without subjecting animals to unnecessary tests or discomfort[40]. Limitations when synthetic images are used: • Realism Constraints: Although synthetic image generation has made significant advancements, the resulting images may still exhibit subtle artefacts or unrealistic features. These imperfections could lead to a degradation in model performance when applied to real-world scenarios, as the models may struggle to generalise from synthetic to actual clinical data [39]. • Bias Propagation: Synthetic datasets, while generated artificially, can inadvertently carry over biases from the real-world data used in the finetuning process. This issue may limit the generalizability of models trained exclusively on synthetic data, as the diversity and complexity of real-world cases might not be fully captured [36]. • Necessity for Clinical Validation: Despite their utility, synthetic images require thorough validation to ensure their clinical relevance. Without rigorous validation processes, models trained on synthetic data might underperform when deployed in real-world clinical settings, especially when tasked with recognising nuanced or rare conditions [40]. 3.2 Model Architecture and Training 3.2.1 U-Net The U-Net architecture is widely regarded as a robust model for image segmentation, particularly in the field of biomedical image analysis. U-Net excels at pixel-wise classification tasks. Its architecture consists of two symmetric parts: an encoder (contracting path) and a decoder (expanding path), forming a U-shaped structure that allows the network to both extract features and reconstruct spatial details at the pixel level. The encoder in U-Net serves to downsample the input image through a series of convolutional and max-pooling layers, extracting increasingly abstract features. Each convolutional block consists of two 3x3 convolutional layers followed by a ReLU activation function. The downsampling occurs through max-pooling layers, which reduce the spatial resolution by a factor of two at each step. This progression allows the U-Net to capture the local features in early layers and more complex global patterns in deeper layers. Visual representation of the U-Net architecture is shown in Figure 7. [56]. At the deepest point in the architecture, the bridge links the encoder and decoder by combining features learned from the encoder’s deepest layer with features from the decoder’s shallowest layer. This design ensures that global and local features are propagated through the network, facilitating accurate segmentation even in cases of complex images with small or subtle regions of interest [70]. The combination of these feature maps is essential for generating detailed and contextually accurate segmentations, particularly in medical applications where high precision is critical. The decoder mirrors the encoder, progressively upsampling the feature maps and restoring the spatial resolution of the input image. The decoder combines the upsampled feature maps with the feature maps from the corresponding encoder layers via skip connections. These skip connections help retain the spatial context and fine details from the encoder, ensuring that the segmentation is accurate at a pixel level [18], [71]. Figure 7. The standard U-Net architecture, where the left encoder part and right decoder part form the letter U. The encoder effectively captures detailed features from the input images, which are then upsampled by the decoder part of the U-Net. The decoder uses transposed convolutions to restore the image to its original size, using skip connections from the corresponding encoder layers to refine the segmentation output with high-resolution details retained [56] Leading Article

RkJQdWJsaXNoZXIy OTc5MDU=