Fast metal artifact correction of multi-contrast MRI by jointly optimizing multi-dimensional sampling pattern with model-based deep learning

Abstract

Metal artifacts caused by metallic implants present a significant challenge in accurate diagnosis using MRI. While various techniques have been developed to mitigate these artifacts, their clinical utility is often limited due to long scan times. Among these techniques, Slice Encoding for Metal Artifact Correction (SEMAC) has gained popularity. In this thesis, we propose an artificial neural network approach to accelerate SEMAC, enabling the suppression of metal artifacts in a shorter scan time without compromising image quality. Building upon previous work on jointly optimizing sampling patterns in MRI reconstruction (J-MoDL), we effectively address metal artifacts by simultaneously optimizing multi-contrast MRI and multi-dimensional sampling patterns. By optimizing sampling positions on both in-plane and SEMAC dimensions, we achieve promising results in fast MRI with enhanced metal artifact correction. Our experiments on patients with degenerative spinal diseases, known for severe metal artifacts, demonstrate the model’s exceptional performance in correcting metal artifacts quantitatively and perceptually. Furthermore, we establish the balanced reconstruction quality across multi-contrast MR images by employing a different denoiser named XNet.