Deep Learning for Image Enhancement and Correction in Magnetic Resonance Imaging-State-of-the-Art and Challenges

Abstract

Magnetic resonance imaging (MRI) provides excellent soft-tissue contrast for clinical diagnoses and research which underpin many recent breakthroughs in medicine and biology. The post-processing of reconstructed MR images is often automated for incorporation into MRI scanners by the manufacturers and increasingly plays a critical role in the final image quality for clinical reporting and interpretation. For image enhancement and correction, the post-processing steps include noise reduction, image artefact correction, and image resolution improvements. With the recent success of deep learning in many research fields, there is great potential to apply deep learning for MR image enhancement, and recent publications have demonstrated promising results. Motivated by the rapidly growing literature in this area, in this review paper, we provide a comprehensive overview of deep learning-based methods for post-processing MR images to enhance image quality and correct image artefacts. We aim to provide researchers in MRI or other research fields, including computer vision and image processing, a literature survey of deep learning approaches for MR image enhancement. We discuss the current limitations of the application of artificial intelligence in MRI and highlight possible directions for future developments. In the era of deep learning, we highlight the importance of a critical appraisal of the explanatory information provided and the generalizability of deep learning algorithms in medical imaging.

Keywords: Artefact correction; Image enhancement; Magnetic resonance imaging; Noise; Post-processing; Super-resolution.

Fig. 1

Overview of the scope of the review paper which focuses on the post-processing steps after image reconstruction and includes MRI artefact correction, noise reduction, and resolution enhancement

Fig. 2

(a) Unet consists of a fully convolutional encoder and decoder interconnected by concatenating feature maps which assists in propagating spatial information to deep network layers. (b) Autoencoder consists of an encoder which maps images to a latent space of reduced dimensionality and a decoder which maps the latent space vector to image space. The dimensionality reduction mitigates random variations in the input while preserving image features necessary for image reconstruction. (c) Generative adversarial network consists of a generator network which produces an estimate of a ground-truth image and a discriminator which attempts to discern between synthesized images and ground truth images. Parameters for each network are updated in an alternating fashion resulting in generator outputs which are indistinguishable from ground truth images from the perspective of the discriminator.

Fig. 3

Motion degraded (left-hand column) and motion-corrected (right-hand column) images highlighting the image quality improvement for a case with a brain tumor. The ringing motion artefacts were removed from the images without degrading the diagnostic image quality [72]

Fig. 4

EPI distortion artefact correction in [80] form three datasets. D: distorted image, U: undistorted images formed using distorted and T1 weighted as input to a 3D Unet

Fig. 5

Comparison of image denoising on a T1-weighted image [30]: (a) noise-free image, (b) noisy image, (c) BM4D, (d) PRI-NLM3D, (e) CNN3D, and (f) RED-WGAN [48]

Fig. 6

Example of a horizontal tear in the body of the lateral meniscus can be identified with the hyperintense double echo in steady-state signal. First column, high-resolution ground-truth; second column, DeepResolve; and third column, tricubic interpolation (TCI). Compared with the Ground-Truth, the DeepResolve image shows considerably less blurring to TCI images [163]