Retrospective MoCo in Liver MRI via U-Net and GAN

Liver MRI plays a crucial role in diagnosing various diseases; however, motion artifacts caused by patient movement during scanning can significantly degrade image quality, leading to misdiagnoses and additional scanning costs. This study explores a deep learning based retrospective motion correction (MoCo) approach using U-Net and Generative Adversarial Networks (GANs) to reduce motion artifacts in liver MRI images. Motion artifactsincluding regular moving motion, ghosting effects, and spiking distortionsare simulated using TorchIO to generate training and validation datasets. The proposed model integrates Fully Convolutional Networks (FCNs), U-Net, and Patch-GAN to enhance feature learning through adversarial training. Additionally, perceptual loss is incorporated to test to improve the models ability to retain high-level details. The performance of the models is evaluated using the Structural Similarity Index (SSIM) to quantify image quality improvements. The study aims to demonstrate that deep learning-based MoCo can enhance liver MRI interpretation accuracy, reduce the need for repeated scans, and improve diagnosing efficiency while minimizing costs associated with motion artifacts.