Optimal MRI Undersampling Patterns for Ultimate Benefit of Medical Vision Tasks

Artem Razumov, Oleg Y. Rogov, Dmitry V. DylovarXiv 2021 · DOI · OpenReview

TL;DR

This paper introduces a novel approach to accelerate MRI scans by optimizing k-space undersampling patterns not for image quality, but for the performance of downstream medical vision tasks like segmentation and classification. Their method, Iterative Gradient Sampling (IGS), learns task-specific undersampling patterns that significantly improve diagnostic accuracy, especially at high acceleration factors where traditional methods struggle.

What Problem Does This Paper Address?

Magnetic Resonance Imaging (MRI) is a powerful diagnostic tool, but its slow scan times are a major bottleneck in clinical practice. Accelerating MRI acquisition is crucial, and a common technique is "compressed sensing," which involves undersampling the raw k-space data and then reconstructing the image. Historically, the optimization of these undersampling patterns has focused solely on maximizing the visual quality of the reconstructed image. However, this paper argues that this focus is misplaced. The ultimate goal of an MRI scan is not just a pretty picture, but to aid in diagnosis. Therefore, the undersampling pattern should be optimized to improve the performance of the specific medical vision task (e.g., tumor segmentation, disease classification) that will be performed on the reconstructed image, even if it means the reconstructed image itself looks less visually appealing.

Methods

The core of this paper is the Iterative Gradient Sampling (IGS) algorithm. Instead of optimizing for image quality metrics like SSIM or PSNR, IGS optimizes for the loss function of a specific downstream medical vision task (e.g., Dice loss for segmentation, cross-entropy for classification).

Here's how it works:

  1. Initialization: The undersampling pattern starts as a null pattern, with a single point in the center (to avoid zero gradients).
  2. Iterative Gradient Calculation: For each iteration, the algorithm:
  3. Takes fully sampled k-space data and undersamples it using the current pattern.
  4. Reconstructs the undersampled image (using a basic zero-filled reconstruction).
  5. Feeds the reconstructed image into a pre-trained medical vision model (e.g., a U-Net for segmentation).
  6. Calculates the loss for that specific task (e.g., Dice loss between the predicted segmentation and the ground truth).
  7. Computes the gradient of this task-specific loss with respect to the undersampling pattern.
  8. Pattern Update: The algorithm identifies the k-space location with the largest negative gradient and "turns on" (sets to 1) that location in the undersampling pattern. This effectively guides the pattern towards sampling the k-space regions that are most informative for the target task.
  9. Iteration: This process is repeated until a desired acceleration factor is reached.

The key innovation is that the undersampling pattern is learned jointly with the downstream task, prioritizing diagnostic utility over visual fidelity.

Datasets

The authors evaluated their method on three widely used medical imaging datasets:

  • ACDC (Automatic Cardiac Diagnosis Challenge): Contains cardiac MRI scans used for segmentation of heart structures.
  • BraTS (Brain Tumor Segmentation): A multi-modal dataset for brain tumor segmentation.
  • FastMRI: A dataset of knee and brain MRI scans, used here for reconstruction and segmentation tasks.

They note that for FastMRI, raw k-space data was available, allowing for direct manipulation of the sampling pattern. For ACDC and BraTS, they used image-level data and simulated k-space undersampling.

Key Insights and Ideas

  • Task-Specific Undersampling is Key: The most significant insight is that optimizing undersampling patterns for the downstream task (segmentation, classification) yields better diagnostic results than optimizing for general image quality. This shifts the paradigm from "making the image look good" to "making the diagnosis accurate."
  • Visual Quality is Secondary: The paper demonstrates that even if the undersampled images appear blurry or have aliasing artifacts, the task-specific optimization can still lead to superior performance in the actual diagnostic task. This is counter-intuitive but crucial for achieving higher acceleration factors.
  • IGS is Universally Applicable: The proposed IGS algorithm is designed to be flexible and can be applied to various medical vision tasks (segmentation, classification, reconstruction) and different MRI modalities.
  • High Acceleration Benefits: The advantages of IGS become most pronounced at high acceleration factors (e.g., ×16 and beyond), where traditional methods produce images of such low quality that diagnostic tasks become unreliable.

Key Results

  • Segmentation Improvement: For the segmentation problem at ×16 acceleration, the IGS-learned patterns achieved up to a 12% improvement in Dice score compared to standard undersampling patterns like FastMRI or center-weighted masks.
  • ACDC Cardiac Segmentation: IGS significantly outperformed other methods, especially at higher acceleration factors, preserving crucial anatomical details for accurate segmentation.
  • BraTS Brain Tumor Segmentation: IGS also showed superior performance, particularly in preserving tumor boundaries and fragments, leading to better Dice scores.
  • Classification: For classification tasks, IGS-learned patterns resulted in more localized and accurate attribution maps, indicating better feature relevance for diagnosis.
  • Reconstruction: While not the primary focus, IGS also showed promise in finding patterns that yield reconstruction quality most similar to fully-sampled images, even with macro-scale aliasing artifacts.

Limitations and Future Work

  • Computational Cost: The iterative nature of the IGS algorithm can be computationally intensive, requiring significant training time.
  • Model Dependency: The learned undersampling pattern is specific to the trained downstream model. If the model architecture changes, the optimal pattern might also change.
  • Artifacts: While diagnostic performance is improved, the reconstructed images can exhibit noticeable aliasing artifacts, especially at high accelerations. The authors mention that these artifacts can often be corrected in post-processing.

Future work suggested by the authors includes:

  • Adapting the IGS strategy to 2D and 3D MRI data more broadly.
  • Exploring domain adaptation techniques for cross-task undersampling patterns.
  • Investigating compound loss functions to optimize multiple medical vision objectives simultaneously.
  • Releasing the code publicly to encourage further research.

Summary updated

Abstract

To accelerate MRI, the field of compressed sensing is traditionally concerned with optimizing the image quality after a partial undersampling of the measurable $\textit{k}$-space. In our work, we propose to change the focus from the quality of the reconstructed image to the quality of the downstream image analysis outcome. Specifically, we propose to optimize the patterns according to how well a sought-after pathology could be detected or localized in the reconstructed images. We find the optimal undersampling patterns in $\textit{k}$-space that maximize target value functions of interest in commonplace medical vision problems (reconstruction, segmentation, and classification) and propose a new iterative gradient sampling routine universally suitable for these tasks. We validate the proposed MRI acceleration paradigm on three classical medical datasets, demonstrating a noticeable improvement of the target metrics at the high acceleration factors (for the segmentation problem at $\times$16 acceleration, we report up to 12% improvement in Dice score over the other undersampling patterns).
BibTeX
@misc{razumov2021optimal,
  author = {Artem Razumov and Oleg Y. Rogov and Dmitry V. Dylov},
  title = {{Optimal MRI Undersampling Patterns for Ultimate Benefit of Medical Vision Tasks}},
  year = {2021},
  eprint = {2108.04914},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2108.04914},
}
Figure