Systems Seminar

Accelerated Medical Imaging: Exploiting Temporal, Spatial, or Parametric Correlation

Prof. Walter Block
Dept. of Biomedical Engineering


Computed tomography (CT), X-ray, and ultrasound imaging have achieved dramatic accelerations in imaging performance over the last year though the addition of greater and greater numbers of detectors. Magnetic resonance imaging (MRI), however, has traditionally been a slow diagnostic imaging modality because spatially encoding the MR signal is not done by a geometric means but is instead acquired in an alternative, Fourier domain. Only recently has MRI developed methods to utilize multiple detectors to offload some of the burden of spatial encoding in the Fourier domain to effectively sample more than one data point at a time. Algorithms used to exploit the redundant information in multiple detectors (coils), termed parallel imaging, become dramatically more complex when used with non-Cartesian acquisition patterns of the MR Fourier space. The first part of the talk describes numerical methods we have developed to utilize and optimize parallel imaging, particularly in non-Cartesian sampling.

In the second part of the talk, the common assumption in medical imaging that all pixels in a temporal or parametric image series are independent and then must be resampled completely for each temporal or parametric index is examined. The University of Wisconsin MR and CT groups has developed a broad class of approximate reconstruction methods, termed HighlY Constrained Back PRojection (HYPR), to utilize the significant redundancy within medical imaging. These methods have superior SNR relative to previous acceleration techniques and often permit transferring the SNR of an entire scan to individual time frames. Often these concepts can be teamed with accelerated non-Cartesian acquisition methods to achieve larger gains. For example, combining an accelerated MR acquisition with a HYPR reconstruction has achieved undersampling factors up to 1000 in selected applications. When applications require a tighter error constraint, we have developed iterative constrained reconstruction methods where the approximate method serves as an initial estimate to speed convergence. Though methods have begun to be developed that merge the HYPR concept with parallel imaging, significant opportunities exist to advance this field though utilizing advanced estimation theory and compressed sensing.

Time and Place: Wed., Nov. 28, at 3:30 pm in 4610.


File "block.shtml" last modified Tue 15 Oct 2019, 01:45 PM, CDT
Web Page Contact: John (dot) Gubner (at) wisc (dot) edu