MR Slice Profile Estimation by Learning to Match Internal Patch Distributions

Shuo Han, Samuel Remedios, Aaron Carass, Michael Schär, Jerry L. Prince

TL;DR

This paper introduces a novel method for estimating the "slice profile" of MRI scans, which describes how the image is blurred in the through-plane direction. By learning to match internal image patch distributions across different orientations, their GAN-based approach can accurately estimate this profile without external data, leading to improved super-resolution results and a way to measure image resolution.

What Problem Does This Paper Address?

Magnetic Resonance Imaging (MRI) often acquires data with lower resolution in the through-plane (slice) direction compared to the in-plane directions. This is a trade-off to reduce scan time and maintain signal-to-noise ratio, which is crucial in clinical settings. To improve image quality and facilitate downstream tasks like registration and segmentation, super-resolution techniques are employed. A key component for many super-resolution algorithms, especially those that learn from paired data, is a "degeneration model" that simulates how a high-resolution (HR) image is converted into a low-resolution (LR) one. In multi-slice 2D MRI, this degeneration is primarily modeled by the slice profile, which acts like a blur kernel.

However, the true slice profile is often unknown and can vary from the theoretically designed one due to factors like pulse sequence imperfections. Existing methods for estimating slice profiles either rely on simulations using the Bloch equation or require imaging physical phantoms, which are not always practical. Some super-resolution algorithms make simplifying assumptions (e.g., a Gaussian profile with FWHM equal to slice separation), which can lead to suboptimal results, especially when the actual slice profile deviates significantly or when there are slice gaps or overlaps. This paper aims to address this gap by proposing a method to accurately estimate the image-specific slice profile directly from the MR data itself, without requiring external calibration or ground truth HR data.

Methods

The core idea of this paper is to leverage the self-similarity of MR images. The authors hypothesize that if an HR image patch is blurred by the correct slice profile and then downsampled, its internal patch distribution will match the distribution of patches extracted from the LR through-plane direction of the original MR volume.

To implement this, they propose a Generative Adversarial Network (GAN) framework. The generator in their GAN is designed to learn a "relative slice profile" (denoted as k). This relative profile is essentially the difference between the through-plane slice profile (pi) and the in-plane point spread function (ph). The generator takes an image patch, blurs it along one in-plane direction using the learned relative slice profile, and then downsamples it. The discriminator is trained to distinguish between these generated LR patches and actual LR patches extracted from the through-plane direction.

By training this GAN, the generator learns to estimate the relative slice profile k that best matches the patch distributions. The authors then derive the actual slice profile pi from k and ph (which can be assumed or estimated separately). The estimation of k is formulated as an internal patch distribution matching problem, making it a self-supervised approach. They also incorporate regularization terms for the slice profile to encourage it to be centered and have zero borders, and use an L2 weight decay for the generator to promote smoothness.

Datasets

The authors use simulated LR images derived from five subjects from the OASIS-3 dataset. These are described as isotropic brain scans with a resolution of 1 mm. They generate simulated LR images by blurring these HR scans with either Gaussian or rect functions of varying FWHMs and then downsampling them with scale factors of 2, 4, and 8.

Key Insights and Ideas

  • Self-Similarity as a Key: The paper hinges on the observation that MR images exhibit similar patterns across different orientations. This self-similarity is exploited to create a learning signal for estimating the slice profile.
  • Relative Slice Profile: Instead of directly learning the full slice profile, learning a "relative" slice profile simplifies the problem and makes it more amenable to GAN-based learning, especially when the in-plane PSF is known or can be approximated.
  • GAN for Distribution Matching: The use of a GAN is a natural fit for this problem, as GANs are powerful tools for learning and matching complex data distributions, which is exactly what's needed to align patch distributions.
  • Slice Profile as a Measurable Property: Beyond its use in super-resolution, the paper demonstrates that their estimated slice profile can be used as a tool to measure the actual through-plane resolution of acquired MR data or super-resolved images.

Key Results

  • The proposed method successfully estimates slice profiles in simulations, achieving good PSNR and SSIM values between the degraded images and the true HR images, even when the estimated slice profile has some error.
  • Incorporating the estimated slice profile into the SMORE super-resolution algorithm led to improved PSNR and SSIM results, particularly when the slice thickness was smaller than the slice separation.
  • The method was shown to be capable of measuring the through-plane resolution of super-resolved images, providing mean FWHM values close to the ground truth in their experiments.

Limitations and Future Work

  • Flexible Shape Limitation: The authors acknowledge that their current generator architecture, designed for smoothness, struggles to accurately learn slice profiles with highly flexible or non-smooth shapes (like rect functions). They suggest that the regularization parameter λw significantly impacts performance, and finding an optimal balance for stable training with flexible shapes is an area for future work.
  • Inconclusive Improvement with SMORE: While improvements were observed, the results with SMORE were not consistently better across all tested scenarios, especially when slice thickness was larger than slice separation (leading to overlaps). They plan to conduct more experiments with SMORE, including longer training times and testing other super-resolution algorithms, to further validate their findings.
  • Future Directions: The authors suggest investigating better ways to regularize the slice profile to allow for more flexible shapes and exploring other super-resolution algorithms to integrate their slice profile estimation method.

Summary updated

BibTeX
@inproceedings{han2021mr,
  author = {Shuo Han and Samuel Remedios and Aaron Carass and Michael Schär and Jerry L. Prince},
  title = {{MR Slice Profile Estimation by Learning to Match Internal Patch Distributions}},
  booktitle = {IPMI},
  year = {2021},
  doi = {10.1007/978-3-030-78191-0\_9},
}

Loading related content...

Figure