CycleGAN

This retrospective study developed a semisupervised deep learning framework using CycleGAN-based style transfer to enable automated kidney segmentation across multiphase contrast-enhanced (MCE) MRI acquisitions. T2-weighted images with manual kidney masks were used to generate anatomically coregistered synthetic images for multiple contrast phases, which then trained Mask R-CNN segmentation models. The method achieved high segmentation accuracy on independent MCE MRI data, with Dice scores ranging from 0.91 to 0.93 across all contrast phases, demonstrating strong cross-phase robustness.