VN October 2025

Vetnuus | October 2025 15 The decoder uses transpose convolutions (deconvolutions) to increase the image resolution and further refine the output segmentation mask [72]. The final output is a pixel-wise segmentation mask, where each pixel is assigned a class label based on the features extracted by the encoder and decoder. U-Net has proven to be effective in tasks that require accurate segmentation with limited training data, making it particularly wellsuited for medical imaging, where annotating large datasets is often challenging [18], [71]. Its symmetrical structure and use of skip connections make it highly capable of handling both large and small objects in an image, and it performs well even when there is significant variation in object size and shape. 3.2.1.1 Transfer learning Transfer learning significantly enhances the capability of deep learning models, particularly in specialised tasks such as medical image segmentation, by leveraging pre-trained neural networks. The U-Net architecture benefits from incorporating robust feature-extracting networks, known as backbones, which are pre-trained on large and diverse datasets like ImageNet [73]. These backbones capture a wide array of features, from simple textures to complex structures, which are crucial for accurate segmentation. One popular choice for a backbone in U-Net is the VGG network [31], known for its simplicity and effectiveness in feature extraction due to its deep architecture of 16 convolutional layers. The typical VGG-16 architecture, as shown in Figure 8. [56], consists of multiple convolutional layers followed by max-pooling layers, which progressively reduce the spatial dimensions while increasing the depth, capturing finer details at each layer. Another powerful backbone is the ResNet18 [74], which incorporates residual connections to allow training of much deeper networks by addressing the vanishing gradient problem. These residual connections act as shortcuts that enable the gradient to flow through the network without diminishing, preserving the strength of the signal (Figure 9. [56]). This architecture is particularly beneficial for U-Net, enhancing its ability to learn from medical images where preservation of spatial hierarchies is critical. The Inception V3 network [75], known for its efficiency in handling multi-scale information, uses modules composed of parallel convolutional layers with varying kernel sizes. This design allows the network to capture features at various scales within the same layer, making it highly effective for tasks that require the detection of objects of different sizes, such as different types of tissues or cells in medical images. The Inception V3 network is shown in Figure 10. [56]. Leading Article Figure 8. The VGG-16 architecture Figure 9. The ResNet18 architecture >>>16

RkJQdWJsaXNoZXIy OTc5MDU=