STRESS: Super-Resolution for Dynamic Fetal MRI Using Self-supervised Learning

Junshen Xu, Esra Abaci Turk, Patricia Ellen Grant, Polina Golland, Elfar Adalsteinsson

TL;DR

This paper introduces STRESS, a novel self-supervised learning framework for enhancing the resolution of dynamic fetal MRI. By cleverly simulating interleaved slice acquisitions and leveraging both spatial and temporal information, STRESS significantly improves image quality, overcoming limitations of traditional methods for capturing rapid fetal motion.

What Problem Does This Paper Address?

Dynamic fetal MRI is crucial for understanding fetal development and function, but it's severely hampered by unpredictable and rapid fetal motion. Fast imaging techniques, while necessary to capture motion, often compromise image quality and resolution. Traditional super-resolution (SR) methods struggle with dynamic fetal MRI, especially when acquiring multiple orientations for oversampling isn't feasible or when high temporal resolution is needed. Furthermore, fetal motion makes it difficult to obtain the high-resolution (HR) ground truth data required for supervised SR methods. This paper tackles the challenge of improving spatial resolution in dynamic fetal MRI without relying on HR ground truth, which is difficult to acquire in this context.

Methods

STRESS employs a self-supervised super-resolution (SSR) approach. The core idea is to generate low-resolution (LR) and high-resolution (HR) image pairs from the already acquired interleaved dynamic fetal MRI data. This is achieved by:

  1. Simulated Interleaved Acquisition: The method simulates an interleaved slice acquisition process on the acquired data. This involves interpolating the existing slices to create a higher-resolution volume and then virtually acquiring slices in an interleaved manner along the desired high-resolution axis. This process yields pairs of LR and HR images.
  2. Spatio-Temporal Training: A super-resolution network is trained using these simulated LR-HR pairs. Crucially, the network exploits both spatial correlations within individual frames and temporal correlations between adjacent frames in the MR time series. This allows it to learn a mapping that enhances resolution.
  3. Optional Blind-Spot Denoising (BDN): Recognizing that fast MRI sequences like EPI can suffer from low signal-to-noise ratio (SNR), STRESS includes an optional blind-spot denoising network. This network is trained to denoise images without seeing the central pixel, preventing it from simply learning an identity function. When enabled, the SR network is trained to reconstruct denoised HR images.

Datasets

The paper evaluates STRESS on two datasets:

  1. CRL Fetal Dataset: This dataset consists of T2-weighted fetal brain MRI volumes with isotropic resolution. To simulate fetal motion, they used existing fetal landmark time series data to generate motion trajectories and create simulated MR scans with different interleave parameters (N₁ = 2, 4, 6). Both noise-free and noisy (Rician noise) data were used.
  2. Fetal EPI Dataset: This dataset comprises in utero fetal EPI (Echo-Planar Imaging) MRI time series. These are naturally acquired with interleaved multislice single-shot sequences. For this dataset, ground truth is not available, so they used the removed slices from the interleaved acquisition as a reference for evaluation.

Key Insights and Ideas

  • Self-supervised learning for dynamic fetal MRI is viable: The paper demonstrates that it's possible to achieve high-quality super-resolution without needing explicit HR ground truth, which is a major bottleneck in fetal MRI.
  • Simulating interleaved acquisitions is key: By cleverly using the existing interleaved acquisition pattern to generate LR-HR pairs, the method bypasses the need for separate high-resolution scans.
  • Exploiting temporal information is crucial: Unlike many SR methods that treat frames independently, STRESS leverages the temporal continuity of dynamic MRI, which is essential for capturing motion and improving reconstruction.
  • Blind-spot denoising is a valuable addition: For low-SNR sequences, the optional BDN significantly improves the performance of the SR network, highlighting the importance of addressing noise in fast MRI.

Key Results

  • Superior Performance: STRESS consistently outperformed competing methods (including interpolation-based methods, SMORE, and others) in terms of peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) on both simulated and real fetal MRI data.
  • Quantitative Improvements: For the CRL dataset, STRESS achieved the highest PSNR and SSIM across different interleave parameters and noise levels. For example, with N₁=4 and Rician noise (σ=3%), STRESS achieved a PSNR of 32.81 and SSIM of 0.9655.
  • Qualitative Improvements: Visual results showed that STRESS produced sharper images with clearer boundaries compared to other methods, particularly in capturing details corrupted by fetal motion.
  • Downstream Task Benefit: When used for fetal keypoint detection (a downstream task), the super-resolved images from STRESS led to improved performance (higher PCK scores) compared to images processed by other SR methods.

Limitations and Future Work

  • Interleave Parameter Dependence: While STRESS performs well, its performance is influenced by the interleave parameter (N₁). The paper notes that as N₁ increases, the challenge of super-resolution becomes greater.
  • Computational Cost: Training deep learning models, especially with large datasets and optional denoising networks, can be computationally intensive.
  • Future Work: The authors suggest that STRESS can be a valuable preprocessing step for various downstream tasks, implying further exploration of its utility in applications like fetal motion tracking and functional analysis. The paper also mentions the potential for further improvements by exploring different network architectures or training strategies.

Summary updated

Abstract

Fetal motion is unpredictable and rapid on the scale of conventional MR scan times. Therefore, dynamic fetal MRI, which aims at capturing fetal motion and dynamics of fetal function, is limited to fast imaging techniques with compromises in image quality and resolution. Super-resolution for dynamic fetal MRI is still a challenge, especially when multi-oriented stacks of image slices for oversampling are not available and high temporal resolution for recording the dynamics of the fetus or placenta is desired. Further, fetal motion makes it difficult to acquire high-resolution images for supervised learning methods. To address this problem, in this work, we propose STRESS (Spatio-Temporal Resolution Enhancement with Simulated Scans), a self-supervised super-resolution framework for dynamic fetal MRI with interleaved slice acquisitions. Our proposed method simulates an interleaved slice acquisition along the high-resolution axis on the originally acquired data to generate pairs of low- and high-resolution images. Then, it trains a super-resolution network by exploiting both spatial and temporal correlations in the MR time series, which is used to enhance the resolution of the original data. Evaluations on both simulated and in utero data show that our proposed method outperforms other self-supervised super-resolution methods and improves image quality, which is beneficial to other downstream tasks and evaluations.
BibTeX
@inproceedings{xu2021stress,
  author = {Junshen Xu and Esra Abaci Turk and Patricia Ellen Grant and Polina Golland and Elfar Adalsteinsson},
  title = {{STRESS: Super-Resolution for Dynamic Fetal MRI Using Self-supervised Learning}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2021},
  doi = {10.1007/978-3-030-87234-2\_19},
}
Figure