Evaluating Generative Adversarial Networks for Virtual Contrast-Enhanced Kidney Segmentation using Res-UNet in Non-Contrast CT Images

This study addresses the challenges of accurately segmenting kidneys in non-contrast CT images; where Convolutional neural networks (CNNs) perform best. The main goal is to investigate how well a Generative Adversarial Network (GAN) performs in producing virtual contrast-enhanced (vCECT) images from non-contrast sources. A GAN model is trained to generate simulated contrast-enhanced images that preserve pertinent information throughout translation by utilizing a dataset of 286 paired non-contrast and contrast-enhanced images. Following, 20 pairs of non-contrast and vCECT pictures are used to assess the performance of a unique segmentation model named Res-U-Net. The Res-U-Net architecture, which combines a U-Net design with residual connections, is optimized to efficiently extract detailed information from original and vCECT images. Image preprocessing methods like Hounsfield windowing with histogram equalization are used to improve the quality of the data before it is fed into the deep neural network. Dice similarity coefficient (DSC) is used to measure Res-U-Net's performance and acts as a benchmark against the most advanced kidney segmentation techniques. ResU-Net's performance is evaluated using the DSC. When compared to state-of-the-art methods for kidney segmentation recognition, the ResU-Net systems with intact connections achieved a DSC value of 0.97 percent for organs recognition and 0.83 percent for segmentation approaches. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.