Optimal MRI Undersampling Patterns for Pathology Localization

Artem Razumov, Oleg Y. Rogov, Dmitry V. DylovMICCAI 2022 · 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 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 long acquisition times are a major bottleneck. While compressed sensing techniques aim to speed up MRI by undersampling k-space data, they traditionally focus on reconstructing images that look good. This paper argues that this focus is misplaced. The true goal of MRI is to aid in diagnosis, which relies on downstream tasks like segmenting tumors or identifying pathologies. Reconstructed images that appear visually appealing might not be optimal for these specific analytical tasks. The authors address this by proposing a method to learn undersampling patterns that directly optimize the performance of these downstream medical vision tasks, rather than just the visual quality of the reconstructed image.

Methods

The core of the paper is the Iterative Gradient Sampling (IGS) algorithm. Instead of optimizing for image quality metrics like SSIM or PSNR, IGS optimizes for task-specific loss functions (e.g., Dice loss for segmentation, cross-entropy for classification). The algorithm starts with a basic undersampling pattern and iteratively "learns" where to sample in k-space by:

  1. Undersampling and Reconstruction: Applying the current undersampling pattern to the fully sampled k-space data and reconstructing the image.
  2. Task-Specific Prediction: Feeding the reconstructed image into a pre-trained medical vision model (e.g., U-Net for segmentation).
  3. Gradient Calculation: Computing the gradient of the task-specific loss function with respect to the undersampling pattern.
  4. Pattern Update: Selecting the k-space location with the largest negative gradient and adding a sample point there (setting the corresponding element in the binary undersampling mask to 1).

This process is repeated, effectively performing a gradient descent on the binary undersampling pattern to maximize the target metric. The key insight is that the "look" of the reconstructed image is secondary; what matters is that the pattern improves the diagnostic outcome.

Datasets

The authors evaluated their method on three well-known medical imaging datasets:

  • ACDC (Automated Cardiac Diagnosis Challenge): Used for cardiac segmentation tasks.
  • BraTS (Brain Tumor Segmentation Challenge): Used for brain tumor segmentation and classification tasks.
  • fastMRI: A benchmark dataset for accelerated MRI, used here for reconstruction and segmentation tasks.

Key Insights and Ideas

  • Task-Specific Undersampling: The most significant insight is that optimizing undersampling patterns for specific downstream tasks (segmentation, classification) yields better diagnostic performance than optimizing for general image quality.
  • "Look" vs. "Function": The paper challenges the conventional wisdom that visually pleasing reconstructions are always best. Their method shows that even if a reconstructed image appears "blurry" or has aliasing artifacts, it can still lead to superior performance in critical diagnostic tasks.
  • High Acceleration Advantage: The IGS approach is particularly beneficial at high acceleration factors (e.g., ×16 and beyond), where traditional methods produce images of such low quality that they become diagnostically useless. IGS maintains a higher level of task performance even in these extreme scenarios.
  • Application-Specific Patterns: The optimal undersampling patterns are not universal; they are application- and mode-specific, meaning a pattern optimized for cardiac segmentation might differ from one optimized for brain tumor classification.

Key Results

  • Segmentation Improvement: For the segmentation problem at ×16 acceleration, the IGS-optimized patterns achieved up to a 12% improvement in Dice score compared to standard undersampling strategies like FastMRI or central masks.
  • Consistent Performance: The method demonstrated consistent improvements across different datasets (ACDC, BraTS) and network architectures (U-Net, U-Net with attention, 3D U-Net).
  • Reconstruction Quality: While the primary focus was on task performance, the IGS approach also showed competitive or improved reconstruction metrics (SSIM, PSNR) compared to standard methods, particularly for the BraTS dataset.
  • Robustness at High Accelerations: Figures 5 and 6 show that as acceleration factors increase, the performance gap between IGS and other methods widens significantly, highlighting its utility for ultrafast MRI.

Limitations and Future Work

  • Computational Cost: The iterative nature of the IGS algorithm can be computationally intensive, requiring significant training time.
  • Generalization: While the paper shows application-specific patterns, further work is needed to explore how well these patterns generalize across different patient populations or scanner variations.
  • Future Directions: The authors suggest several avenues for future research:
  • Adapting the undersampling strategy to 2D and 3D imaging.
  • Applying domain adaptation methods for cross-task undersampling.
  • Investigating compound loss functions to optimize multiple medical vision objectives simultaneously.
  • Exploring the use of these patterns for non-diagnostic applications like surgical planning or radiation therapy planning.

Summary updated

Abstract

We investigate MRI acceleration strategies for the benefit of downstream image analysis tasks. Specifically, we propose to optimize the k-space undersampling patterns according to how well a sought-after pathology could be segmented or localized in the reconstructed images. We study the effect of the proposed paradigm on the segmentation task using two classical labeled medical datasets, and on the task of pathology visualization within the bounding boxes, using the recently released fastMRI+ annotations. We demonstrate a noticeable improvement of the target metrics when the sampling pattern is optimized, e.g., for the segmentation problem at $$\times $$ 16 acceleration, we report up to 12% improvement in Dice score over the other undersampling strategies.
BibTeX
@inproceedings{razumov2022optimal,
  author = {Artem Razumov and Oleg Y. Rogov and Dmitry V. Dylov},
  title = {{Optimal MRI Undersampling Patterns for Pathology Localization}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2022},
  url = {https://openreview.net/forum?id=nCv7qiFOx85},
}
Figure