Explicit Abnormality Extraction for Unsupervised Motion Artifact Reduction in Magnetic Resonance Imaging

TL;DR

MRI motion artifacts degrade diagnostic quality and obtaining paired MA-corrupted/MA-free data is impractical. The paper introduces UNAEN, an unsupervised framework with an explicit artifact extractor $G_e$ and a reconstructor $G_r$ that operate on unpaired MA-free and MA-corrupted images, producing MA-reduced images $x = x^a - G_e(x^a)$ and cycle-consistent reconstructions $ar{x}^a$. Using an RCAN-based backbone, adversarial and cycle-consistency losses are combined to train $G_e$ and $G_r$, while a forward/backward discriminator setup enforces realism. Empirically, UNAEN outperforms state-of-the-art unsupervised MAR methods on fastMRI and BraTS datasets, with qualitative results showing sharper structures and fewer residual artifacts, albeit with higher computational cost. The approach reduces reliance on paired data and has potential for broader clinical impact and extension to other image-quality improvements beyond motion artifact reduction.

Abstract

Motion artifacts compromise the quality of magnetic resonance imaging (MRI) and pose challenges to achieving diagnostic outcomes and image-guided therapies. In recent years, supervised deep learning approaches have emerged as successful solutions for motion artifact reduction (MAR). One disadvantage of these methods is their dependency on acquiring paired sets of motion artifact-corrupted (MA-corrupted) and motion artifact-free (MA-free) MR images for training purposes. Obtaining such image pairs is difficult and therefore limits the application of supervised training. In this paper, we propose a novel UNsupervised Abnormality Extraction Network (UNAEN) to alleviate this problem. Our network is capable of working with unpaired MA-corrupted and MA-free images. It converts the MA-corrupted images to MA-reduced images by extracting abnormalities from the MA-corrupted images using a proposed artifact extractor, which intercepts the residual artifact maps from the MA-corrupted MR images explicitly, and a reconstructor to restore the original input from the MA-reduced images. The performance of UNAEN was assessed by experimenting with various publicly available MRI datasets and comparing them with state-of-the-art methods. The quantitative evaluation demonstrates the superiority of UNAEN over alternative MAR methods and visually exhibits fewer residual artifacts. Our results substantiate the potential of UNAEN as a promising solution applicable in real-world clinical environments, with the capability to enhance diagnostic accuracy and facilitate image-guided therapies. Our codes are publicly available at https://github.com/YuSheng-Zhou/UNAEN.

Figures (6)

• Figure 1: The schematic of UNAEN. The proposed method contains a forward module (blue block), which reduces the artifacts of MA-corrupted images, and a backward module (yellow block), which restore the MA-corrupted images based on the forward module outputs. MA-reduced images can be obtained by explicitly extracting artifacts of MA-corrupted images using $G_e$ and deleting them directly by the forward module, where $D_f$ compares the MA-reduced images with the MA-free images to identify whether the artifact removal is successful. The backward module converts the MA-reduced images to the original input by $G_r$ and $D_b$ is used to check whether $G_r$ is restored successfully. There is a cycle consistency between the MA-corrupted images and the restored MA-corrupted images.
• Figure 2: The detailed structures of generator and discriminator. The generator adopts the RCAN backbone with a depth of 5 residual groups (RG) and a long skip connection, and the discriminator is a VGG network.
• Figure 3: Illustration of motion artifact extractions with varying degrees of severity in each row. MA Maps column (c) denotes the error between MA-free images (Ground Truth in column (a)) and corresponding MA-corrupted images (column (b)). The proposed network was trained to extract residual artifact maps (Extracted MA MAP in column (d)) explicitly. The MA-reduced column (e) displays the restored images after MA reduction. Highly consistent patterns can be observed between the MA Map column and extracted MA Map column, revealing the effectiveness of the explicit MA extraction.
• Figure 4: Comparison of the qualitative performance of UNAEN and other unsupervised models on the fastMRI brain dataset. There are visualizations of the artifact reduction results with varying degrees of artifact severity and corresponding error heat maps showing the difference between ground truth and each result.
• Figure 5: Comparison of the qualitative performance of UNAEN and other unsupervised models on the BraTS dataset. This visualized the artifact reduction results with varying degrees of artifact severity and corresponding error heat maps showing the difference between ground truth and each result.