McAteer, Joshua
(2024)
Optimising MRI Sampling and Reconstruction to Maximise Image Information Quality and Minimise Measurement Error.
PhD thesis, University of Nottingham.
Abstract
This thesis covers the development of optimised k-space sampling compressed sensing MR imaging. Methods and analysis tools were developed and assessed to improve image quality in compressed sensing accelerated MR imaging. The advantages and limitations of optimised k-space sampling are discussed as well as conventional methods of image acquisition acceleration.
The basic phenomena of nuclear magnetic resonance and MRI were described including the source of contrast. Then image formation was explained as well as artifacts that occur in MR imaging.
Automated image quality assessment is vital to image acquisition optimisation and was explored in this work. This included the evaluation of a large number of image quality metrics and the development of several new metrics. These metrics were compared against one-another as well as against common MR imaging aberrations, such as Rician noise and compressed sensing artifacts. Blind image quality metrics performed poorly, since they make assumptions about the unaberrated image which may not be true. Some of the referenced image quality metrics also performed poorly, including PSNR, CORR-PY, and PSNR-HVS-M. The most useful image quality metrics were CORR, HP-CORR, MI, VIF, SDRES, B-SDRES, RMSE, HP-RMSE, and PIQE (which was the best of the blind metrics). However, the relative sensitivity of each metric to each aberration type varies greatly, and so it is important to select an image quality metric that measures the aspect of image quality that is most important for the particular investigation. In cases where a quantitative measurement is the main/only output that measurement itself can be used as an information quality metric (by measuring the error) which is a more direct method of assessing image suitability.
Conventional methods of image acceleration were investigated including partial Fourier, parallel imaging, and compressed sensing. These methods have different use cases and requirements and so the selection of the method is dependent on setup and imaging goals. Partial Fourier cannot accelerate images by more than 50% (typical acceleration is slightly less than 50%). In addition the image quality is quite poor compared with other methods. Parallel imaging requires specific hardware (multiple receive coils in the phase encode direction) and is able to produce high quality
images. However, there is a noise amplification effect that typically occurs near the centre of the image. Compressed sensing does not require specific hardware, but does require software support that is not universal (but is becoming more common). High acceleration factors are possible as are high quality images. However, the non-linear properties of the imaging require experience to interpret. In addition, compressed sensing can be used along with parallel imaging to take advantage of both methods. This allows for higher quality image or faster acquisition than either alone.
In order to improve compressed sensing imaging optimised k space sampling was investigated along with optimised regularisation of the reconstruction of compressed sensing images. The kspace sampling optimisation methods used training data from prior scans to develop efficient k-space sampling schemes that sampled novel data significantly better that typical heuristic rules. Two algorithms were developed for this: power-law optimisation and line-wise optimisation. Line-wise optimisation was slower and more rigorous and so yielded more efficient sampling and better image quality. These k-space sampling schemes were found to be highly tolerant to changes in image structure, such as warping and rotation, as well as reasonably tolerant to changes in image content. In addition, the methods described are simple, clear, and understandable and so can be used without concern of spurious detail hallucination that is possible with some deep learning methods.
The optimised k-space sampling schemes were applied to problems that were in situ and in vivo. American College of Radiology phantoms were imaged, as well as three organs: The brain, stomach, and lungs. This demonstrated the effectiveness of the methods under a variety of real world conditions. These also required different end points, including the ability to detect a compact region, direct image quality assessment, automatic detection of tissue boundary, and automatic segmentation using a deep learning algorithm.
The methods in the work significantly improved image quality in compressed sensing imaging. This allows for better quality images to be acquired in the same time (which may improve diagnostic accuracy) or similar quality images to be acquired in a shorter time (that could be used to improve patient throughput).
| Item Type: |
Thesis (University of Nottingham only)
(PhD)
|
| Supervisors: |
Gowland, Penny
Glover, Paul |
| Keywords: |
magnetic resonance, MRI, Compressed sensing, accelerated imaging, k-space sampling optimisation, image quality analysis, image quality metrics, information quality. |
| Subjects: |
Q Science > QC Physics > QC501 Electricity and magnetism |
| Faculties/Schools: |
UK Campuses > Faculty of Science > School of Physics and Astronomy |
| Item ID: |
77030 |
| Depositing User: |
McAteer, Mr Joshua
|
| Date Deposited: |
23 Jul 2024 04:40 |
| Last Modified: |
23 Jul 2024 04:40 |
| URI: |
https://eprints.nottingham.ac.uk/id/eprint/77030 |
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
