Many different iterative reconstruction algorithms are used in tomography. With each, there are a number of free parameters. The user must decide how many iterations to perform and what values to use for certain smoothing or regularizing parameters. Most users of such algorithms simply adjust the parameters for a visuallly pleasing image, but we have been able to optimize them in terms of task performance. Realistic data sets simulation different organ systems with and without tumors have been created and then reconstructed by the algorithm under test with various parameter setting. Literally thousands of images are simulated for each study. The images are presented to human observers to determine how well they can detect the tumor, and various model observers are also applied. The results are then compared to see how well the models predict human performance.
Some of the many algorithms that have been studied this way include the Landweber algorithm, a maximum-likelihood algorithm known as expectation-maximization (EM), filtered back-projection, and various algorithms that account for attenuation of radiation in the body. Three different students from the Applied Mathematics Program (Edward Soares, Craig Abbey and Brandon Gallas) have been involved in developing the models and performing the psychophysical studies.
The general outcome of these studies is that clearly optimal parameter setting are found and that the current human models are excellent predictors of actual human performance.