Vetnews | Oktober 2025 16 « BACK TO CONTENTS Application of computer vision methods in veterinary ophthalmology..... <<< 15 Improved U-Net with Backbone Integration: The improved U-Net model incorporates these backbone networks into its architecture, where each encoder block is an instance of either VGG-16, ResNet18, or Inception V3. This integration allows the U-Net to utilise the advanced feature extraction capabilities of these networks, enhancing its performance significantly. The decoder of the improved U-Net model up-samples the feature maps and combines them with the extracted features at corresponding levels from the encoder, optimising the segmentation accuracy as seen in Figure 11. [56]. 3.2.1.2 U-Net Algorithm The training of the U-Net model involves an iterative process of adjusting network weights to minimise the loss between the predicted segmentation masks and the ground truth. The algorithm begins by initialising the weights of both the encoder and decoder networks randomly. The model is then trained using a dataset consisting of training images and corresponding ground truth segmentation masks. For each training image, the following steps are performed in the training loop: 1. The encoder network encodes the input image, extracting a set of feature maps at different levels of abstraction. 2. The feature maps from the encoder are concatenated with the corresponding feature maps from the bridge, which combines contextual information from both high- and low-level layers of the network. 3. The concatenated feature maps are then passed through the decoder network, which progressively upsamples the feature maps and reconstructs the spatial resolution to generate a predicted segmentation mask. 4. The loss between the predicted segmentation mask and the ground truth segmentation mask is calculated using the selected loss function (such as Dice loss or cross-entropy loss). 5. The encoder and decoder network weights are updated using an optimisation algorithm (e.g., stochastic gradient descent or Adam) to minimise the loss. The algorithm continues to iterate through the training images, updating the network weights after each image and its Figure 10. The Inception V3 architecture Figure 11. Schematic of the enhanced U-Net architecture with a backbone in both encoder and decoder blocks. The encoder blocks (EB) downscale the input while extracting features, and the decoder blocks (DB) upscale these features to reconstruct the output image U-Net Algorithm Input: Training images and corresponding ground truth segmentation masks Output: trained U-Net model Initialise encoder and decoder network weights randomly. Training Loop For each training image and ground truth segmentation mask: Encode the image using the encoder network. Concatenate encoder feature maps with bridge feature maps Decode concatenated feature maps using a decoder network to generate a predicted segmentation mask. Calculate the loss between the predicted and ground truth segmentation masks Update encoder and decoder network weights using an optimisation algorithm. Convergence Check If loss converges or reaches a satisfactory level: Exit training loop Else: Continue training loop Leading Article Algorithm 1: U-Net Algorithm for Segmentation [56]
RkJQdWJsaXNoZXIy OTc5MDU=