[Paper Review] Data augmentation for deep learning based accelerated MRI reconstruction with limited data

Zalan Fabian, Reinhard Heckel+1 more|arXiv (Cornell University)|Jun 28, 2021

Medical Imaging Techniques and ApplicationsMedicine22 citations

TL;DR

This paper proposes MRAugment, a physics-informed data augmentation pipeline for deep learning-based accelerated MRI reconstruction that leverages MRI acquisition invariances to reduce overfitting in low-data regimes. By applying spatial transformations and noise-preserving augmentations to k-space data, MRAugment improves generalization and robustness, achieving state-of-the-art performance with significantly less training data—matching or exceeding SOTA results while reducing data requirements by up to 98% in some settings.

ABSTRACT

Deep neural networks have emerged as very successful tools for image restoration and reconstruction tasks. These networks are often trained end-to-end to directly reconstruct an image from a noisy or corrupted measurement of that image. To achieve state-of-the-art performance, training on large and diverse sets of images is considered critical. However, it is often difficult and/or expensive to collect large amounts of training images. Inspired by the success of Data Augmentation (DA) for classification problems, in this paper, we propose a pipeline for data augmentation for accelerated MRI reconstruction and study its effectiveness at reducing the required training data in a variety of settings. Our DA pipeline, MRAugment, is specifically designed to utilize the invariances present in medical imaging measurements as naive DA strategies that neglect the physics of the problem fail. Through extensive studies on multiple datasets we demonstrate that in the low-data regime DA prevents overfitting and can match or even surpass the state of the art while using significantly fewer training data, whereas in the high-data regime it has diminishing returns. Furthermore, our findings show that DA can improve the robustness of the model against various shifts in the test distribution.

Motivation & Objective

• Address the challenge of limited training data in deep learning-based MRI reconstruction, where large datasets are costly and time-consuming to collect.
• Overcome the limitations of naive data augmentation in inverse problems like MRI reconstruction, where standard image-level transformations distort measurement consistency.
• Develop a data augmentation pipeline tailored to the physical constraints of MRI acquisition to improve model generalization and robustness.
• Demonstrate that physics-aware data augmentation can significantly reduce data requirements while maintaining or improving reconstruction quality.
• Investigate the impact of data augmentation on model robustness to distribution shifts and hallucination in medical image reconstruction.

Proposed method

• Propose MRAugment, a data augmentation pipeline that applies spatial transformations (e.g., rotations, flips) to k-space data while preserving the underlying physics of MRI measurements.
• Ensure that augmented data maintains consistent noise distribution by applying transformations to the image domain before forward k-space projection, avoiding mismatched noise in training and validation.
• Use an exponential scheduling strategy for augmentation probability to gradually increase the strength of augmentation during training, improving generalization.
• Apply rigid body transformations (rotation, flip, shift) to images prior to k-space sampling, ensuring that the resulting measurements remain consistent with the acquisition model.
• Integrate the augmentation pipeline into a U-Net-based deep learning reconstruction model, training end-to-end with standard L2 loss.
• Validate the method across multiple datasets (brain and knee MRI) and under various undersampling schemes and scanner shifts.

Experimental results

Research questions

• RQ1 Can physics-informed data augmentation reduce the required amount of training data for deep learning-based MRI reconstruction without sacrificing performance?
• RQ2 How does MRAugment compare to naive data augmentation techniques in terms of reconstruction quality and generalization?
• RQ3 To what extent does data augmentation improve model robustness to distribution shifts, such as scanner or field strength differences?
• RQ4 Does data augmentation reduce hallucinated features in reconstructed MRI images, especially in low-data regimes?
• RQ5 Can data augmentation improve generalization when training data is scarce or subject to rapid distribution shifts?

Key findings

• MRAugment significantly reduces overfitting in low-data regimes, achieving state-of-the-art reconstruction quality with as little as 2% of the full training dataset.
• On the fastMRI brain dataset, MRAugment improved SSIM from 0.8646 (no DA) to 0.9049 when training on only 2% of data, outperforming non-augmented models.
• Models trained with MRAugment showed superior robustness to scanner shifts: SSIM improved from 0.8241 to 0.8551 when testing on 3T data after training on 1.5T data.
• Naive data augmentation—applied directly to images without considering noise distribution—degraded performance, causing high-frequency artifacts and lower image quality due to distribution mismatch.
• The use of MRAugment reduced hallucinated features in reconstructions, particularly in low-data settings, as evidenced by visual comparisons and quantitative SSIM gains.
• In high-data regimes, the benefits of data augmentation diminish, indicating that MRAugment is most effective when data is scarce.