SPICER: Self-Supervised Learning for MRI with

Automatic Coil Sensitivity Estimation and Reconstruction

Magnetic Resonance Medicine (MRM) 2024

The FIRST Self-Supervised Method for Automatic Coil Sensitivity Calibration in MRI

Yuyang Hu*1, Weijie Gan*1, Chunwei Ying2, Tongyao Wang3,
Cihat Eldeniz3, Jiaming Liu1, Yasheng Chen4 , Hongyu An2,3,4, and Ulugbek S. Kamilov1

1Computational Imaging Group (CIG), Washington University in St. Louis, St. Louis, MO, USA
2Mallinckrodt Institute of Radiology, Washington University in St. Louis, St. Louis, MO, USA
3Department of Biomedical Engineering,Washington University in St. Louis, St. Louis, MO, USA
4Department of Neurology, Washington University in St. Louis, St. Louis, MO, USA

Yuyang Hu and Weijie Gan contributed equally to this project.

Preprint MRM Code

Abstract

Deep model-based architectures (DMBAs) integrating physical measurement models and learned image regularizers are widely used in parallel magnetic resonance imaging (PMRI). Traditional DMBAs for PMRI rely on pre-estimated coil sensitivity maps (CSMs) as a component of the measurement model. However, estimation of accurate CSMs is a challenging problem when measurements are highly undersampled. Additionally, traditional training of DMBAs requires high-quality groundtruth images, limiting their use in applications where groundtruth is difficult to obtain. This paper addresses these issues by presenting SPICE as a new method that integrates self-supervised learning and automatic coil sensitivity estimation. Instead of using pre-estimated CSMs, SPICE simultaneously reconstructs accurate MR images and estimates high-quality CSMs. SPICE also enables learning from undersampled noisy measurements without any groundtruth. We validate SPICE on experimentally collected data, showing that it can achieve state-of-the-art performance in highly accelerated data acquisition settings (up to 10×).

Banner

Video 1: Illustration of SPICE reconstructed 3D images from real data under 10x acceleration. The video also illustrates the performance of two reference methods, GRAPPA and SSDU.

Model

Figure 1: The SPICE method consists of a DMBA-based MRI reconstruction module and a coil sensitivity estimation module that map multicoil undersampled measurements to a single high-quality image and a set of coil sensitivity maps, respectively. The network is trained directly on raw k-space measurements where the input and the target measurement correspond to a pair of undersampled measurements from the same object.

Validation on Experimentally-Collected 3D MRI Data

Banner

Figure 2: The first row shows the cartesian equispaced sampling masks P used in the experimental validation: R = 4, 6, 8 and 10, with corresponding ACS lines = 24, 24, 8 and 5. The second row shows the zero-filled images of the include the PSNR/SSIM values with respect to the groundtruth.

Banner

Figure 3: Illustration of SPICE reconstructed images compared against several baseline methods.

Banner

Figure 4: Illustration of SPICE reconstructed images compared against several ablation methods.

Banner

Figure 5: Evaluation of CSMs estimated by SPICE at various acceleration rates. The third rows shows the sensitivity map of the 10th coil estimated by SPICE and ESPIRiT. The first row shows the results from using these CSMs within TV reconstruction at acceleration rate R = 4.

Paper

Bibtex

@article{Hu2024SPICER, title={SPICER: Self-Supervised Learning for MRI with Automatic Coil Sensitivity Estimation and Reconstruction}, author={Hu, Yuyang and Gan, Weijie and Ying, Chunwei and Wang, Tongyao and Eldeniz, Cihat and Liu, Jiaming and Chen, Yasheng and An, Hongyu and Kamilov, Ulugbek S.}, journal={Magn. Reson. Med.}, note={DOI:10.1002/mrm.30121}, year={2024} }