Performance-based assessment of reconstructed images
- Los Alamos National Laboratory
During the early 90s, I engaged in a productive and enjoyable collaboration with Robert Wagner and his colleague, Kyle Myers. We explored the ramifications of the principle that tbe quality of an image should be assessed on the basis of how well it facilitates the performance of appropriate visual tasks. We applied this principle to algorithms used to reconstruct scenes from incomplete and/or noisy projection data. For binary visual tasks, we used both the conventional disk detection and a new challenging task, inspired by the Rayleigh resolution criterion, of deciding whether an object was a blurred version of two dots or a bar. The results of human and machine observer tests were summarized with the detectability index based on the area under the ROC curve. We investigated a variety of reconstruction algorithms, including ART, with and without a nonnegativity constraint, and the MEMSYS3 algorithm. We concluded that the performance of the Raleigh task was optimized when the strength of the prior was near MEMSYS's default 'classic' value for both human and machine observers. A notable result was that the most-often-used metric of rms error in the reconstruction was not necessarily indicative of the value of a reconstructed image for the purpose of performing visual tasks.
- Research Organization:
- Los Alamos National Laboratory (LANL), Los Alamos, NM (United States)
- Sponsoring Organization:
- USDOE
- DOE Contract Number:
- AC52-06NA25396
- OSTI ID:
- 956370
- Report Number(s):
- LA-UR-09-00652; LA-UR-09-652; TRN: US201013%%81
- Resource Relation:
- Journal Volume: 7263; Conference: SPIE Medical Imaging Symposium ; February 7, 2009 ; Orlando, FL
- Country of Publication:
- United States
- Language:
- English
Similar Records
Method to evaluate image-recovery algorithms based on task performance
Optimization of the constrained algebraic reconstruction technique for the performance of a variety of visual tasks