Speaker
Description
Magnetic resonance imaging (MRI) is a non-invasive medical imaging technique that provides detailed information about the internal structure and function of the human body. MRI signals are acquired and stored in k-space, which is the spatial frequency domain representation of these signals. The signals are then reconstructed into MR images using linear operators, such as Fourier Transform. However, acquiring fully-sampled k-space data is time-consuming, resulting in patient discomfort and clinical inefficiency. Consequently, under-sampled k-space data is acquired to decrease scan time, despite of causing an ill-posed problem$^{[1]}$. This ill-posed problem leads to instability in the reconstruction process.
To tackle the ill-posed problem in MRI, additional constraints are employed to isolate a unique solution. Several advanced techniques exploiting additional constraints have been developed to improve stability and image quality. In this abstract, we provide an overview of three prominent reconstruction methods: Sensitivity Encoding$^{[2]}$ (SENSE), Compressed Sensing$^{[3]}$ (CS), and Deep Learning (DL)-based$^{[4-7]}$.
-
The sensitivity profile constraint
SENSE reconstruction$^{[2]}$ is a parallel imaging technique in MRI that accelerates data acquisition by using multiple receiver coils with unique sensitivity profiles. The method combines reconstructed images of under-sampled k-space from each coil using these sensitivity profiles by solving a linear system to reconstruct the full image. Therefore, estimating accurate sensitivity profiles is the main key to SENSE reconstruction$^{[8]}$. By using coil sensitivity information, SENSE is capable of accurately reconstructing images with fewer artifacts and superior spatial resolution, even from under-sampled data. -
The sparsity constraint
CS reconstruction$^{[3]}$ is another MRI reconstruction technique that capitalizes on the sparsity of MR images in certain transform domains, such as wavelet. The conditions for successful CS include signal sparsity and incoherence in the sampling pattern. Signal sparsity enables to represent MR images with a few coefficients, helping to reduce the degrees of freedom in the reconstruction problem. Incoherence induces noise-like artifacts which can be removed by non-linear reconstruction. Therefore, the method uses fewer samples to reconstruct images, which can significantly accelerate MRI scans. -
The data-driven constraint
DL-based reconstruction is a more recent approach that employs convolutional neural networks (CNNs) to learn a mapping function between under-sampled data and the corresponding ground truth. This learning process, driven by data, optimizes non-convex objective functions to capture complex patterns and features inherent in the data. Since it is not always possible to acquire fully-sampled ground truth or a large amount of training dataset in MR images, it has become a very active research field to train the networks without ground truth$^{[5,6]}$ or a large dataset$^{[6,7]}$. Given the well-trained model, DL-based reconstructions have demonstrated superior performance compared to traditional methods in terms of both image quality and computational efficiency.
In conclusion, MRI reconstruction is a crucial step for obtaining high-quality MR images from under-sampled k-space data. Techniques such as SENSE, CS, and DL-based reconstruction employ additional constraints to address the ill-posed problem, enhancing image quality and stability. These advanced methods, each with unique advantages, have demonstrated significant improvements in reconstruction performance, paving the way for more efficient and accurate MR imaging.
References
[1] LIU, X. F., XU, W. L., & YE, X. Z. (2009). THE ILL-POSED PROBLEM AND REGULARIZATION IN PARALLEL MAGNETIC RESONANCE IMAGING. 3RD INTERNATIONAL CONFERENCE ON BIOINFORMATICS AND BIOMEDICAL ENGINEERING, ICBBE 2009. HTTPS://DOI.ORG/10.1109/ICBBE.2009.5163622
[2] PRUESSMANN, K. P., WEIGER, M., SCHEIDEGGER, M. B., & BOESIGER, P. (1999). SENSE: SENSITIVITY ENCODING FOR FAST MRI. MAGNETIC RESONANCE IN MEDICINE, 42(5), 952–962. HTTPS://DOI.ORG/10.1002/(SICI)1522-2594(199911)42:5<952::AID-MRM16>3.0.CO;2-S
[3] LUSTIG, M., DONOHO, D., & PAULY, J. M. (2007). SPARSE MRI: THE APPLICATION OF COMPRESSED SENSING FOR RAPID MR IMAGING. MAGNETIC RESONANCE IN MEDICINE, 58(6), 1182–1195. HTTPS://DOI.ORG/10.1002/MRM.21391
[4] SRIRAM, A. ET AL. (2020). END-TO-END VARIATIONAL NETWORKS FOR ACCELERATED MRI RECONSTRUCTION. IN: , ET AL. MEDICAL IMAGE COMPUTING AND COMPUTER ASSISTED INTERVENTION – MICCAI 2020. MICCAI 2020. LECTURE NOTES IN COMPUTER SCIENCE(), VOL 12262. SPRINGER, CHAM. HTTPS://DOI.ORG/10.1007/978-3-030-59713-9_7
[5] YAMAN, B., HOSSEINI, S. A. H., MOELLER, S., ELLERMANN, J., UĞURBIL, K., & AKÇAKAYA, M. (2020). SELF-SUPERVISED LEARNING OF PHYSICS-GUIDED RECONSTRUCTION NEURAL NETWORKS WITHOUT FULLY SAMPLED REFERENCE DATA. MAGNETIC RESONANCE IN MEDICINE, 84(6), 3172–3191. HTTPS://DOI.ORG/10.1002/MRM.28378
[6] YAMAN, B., HOSSEINI, S. A. H., & AKÇAKAYA, M. (2021). ZERO-SHOT SELF-SUPERVISED LEARNING FOR MRI RECONSTRUCTION. HTTP://ARXIV.ORG/ABS/2102.07737
[7] DARESTANI, M. Z., LIU, J., & HECKEL, R. (2022). TEST-TIME TRAINING CAN CLOSE THE NATURAL DISTRIBUTION SHIFT PERFORMANCE GAP IN DEEP LEARNING BASED COMPRESSED SENSING. HTTP://ARXIV.ORG/ABS/2204.07204
[8] UECKER, M., LAI, P., MURPHY, M. J., VIRTUE, P., ELAD, M., PAULY, J. M., VASANAWALA, S. S., & LUSTIG, M. (2014). ESPIRIT - AN EIGENVALUE APPROACH TO AUTOCALIBRATING PARALLEL MRI: WHERE SENSE MEETS GRAPPA. MAGNETIC RESONANCE IN MEDICINE, 71(3), 990–1001. HTTPS://DOI.ORG/10.1002/MRM.24751
| Keywords | Deep Learning, MRI, Image Processing, AI, Image Reconstruction |
|---|
