[Paper Review] MRI Banding Removal via Adversarial Training

Aaron Defazio, Tullie Murrell+1 more|arXiv (Cornell University)|Jan 23, 2020

Medical Imaging Techniques and Applications20 references8 citations

TL;DR

This paper proposes an adversarial training approach to remove banding artifacts in deep learning-based MRI reconstructions without requiring human annotations or additional computation. By training a discriminator to detect banding patterns aligned with the phase-encoding direction, the method significantly reduces visible artifacts while preserving diagnostic detail, outperforming baseline models in a blinded radiologist evaluation.

ABSTRACT

MRI images reconstructed from sub-sampled Cartesian data using deep learning techniques often show a characteristic banding (sometimes described as streaking), which is particularly strong in low signal-to-noise regions of the reconstructed image. In this work, we propose the use of an adversarial loss that penalizes banding structures without requiring any human annotation. Our technique greatly reduces the appearance of banding, without requiring any additional computation or post-processing at reconstruction time. We report the results of a blind comparison against a strong baseline by a group of expert evaluators (board-certified radiologists), where our approach is ranked superior at banding removal with no statistically significant loss of detail.

Motivation & Objective

• To address the major clinical barrier of banding artifacts in deep learning-based MRI reconstructions.
• To develop a method that reduces banding without requiring human-annotated data or post-processing.
• To maintain image detail while improving visual quality for radiologists.
• To create a plug-in solution compatible with any supervised deep learning MRI reconstruction framework.

Proposed method

• Introduce an orientation adversary that discriminates between reconstructed images based on the presence of banding artifacts aligned with the phase-encoding direction.
• Train the adversary using only reconstructed images, not ground-truth images, to avoid introducing phantom anatomical details.
• Integrate the adversarial loss into the training objective of a standard supervised MRI reconstruction model.
• Use a masking procedure to focus the adversarial loss on regions where banding is most visible, such as low-contrast or high-noise areas.
• Train the generator and discriminator jointly via standard GAN optimization, with the generator learning to produce banding-free outputs.
• Apply the method to state-of-the-art MRI reconstruction models without modifying their architecture or inference pipeline.

Experimental results

Research questions

• RQ1 Can adversarial training effectively reduce banding artifacts in deep learning-based MRI reconstructions without human annotations?
• RQ2 Does the proposed method preserve fine anatomical details while removing banding?
• RQ3 Can the orientation adversary detect banding patterns without relying on ground-truth images?
• RQ4 Is the method compatible with existing supervised MRI reconstruction frameworks?
• RQ5 Does the method improve radiologist perception of image quality in a blinded evaluation?

Key findings

• In a blinded evaluation by board-certified radiologists, the proposed method was ranked superior to the baseline in banding removal with no statistically significant loss of detail.
• Radiologists reported banding presence in 72.5% of baseline reconstructions (95% CI: 0.62–0.82), while the adversarial method reduced this to a median of 20% across evaluators.
• The method achieved superior performance without requiring additional computation or post-processing at inference time.
• The orientation adversary successfully detected banding patterns using only reconstructed images, avoiding the risk of introducing phantom anatomical structures.
• The approach is universally applicable to any supervised deep learning MRI reconstruction model, as it operates as a plug-in training loss.
• A reference implementation is publicly available in the supplementary material for reproducibility and integration.