Program Information
A Task-Based Framework for Assessing the Performance of CT Metal Artifact Reduction Algorithms
J. Y. Vaishnav1, 2 , L. Jiang1*, R. Zeng2 , K. J. Myers2 , (1) Office of In Vitro Diagnostics and Radiological Health, US Food & Drug Administration, Silver Spring, MD, (2) Office of Science and Engineering Laboratories, US Food & Drug Administration, Silver Spring, MD
Presentations
TU-H-CAMPUS-IT-1 (Tuesday, August 1, 2017) 4:30 PM - 5:30 PM Room: Imaging ePoster Theater
Purpose: Although several CT metal artifact reduction (MAR) algorithms are commercially available, no standard method exists for assessing their performance. We explore a phantom-based assessment framework based on model observer performance at a signal detection or localization task.
Methods: We design a numerical head phantom with metal implants. In order to incorporate an element of randomness, the phantom includes a rotatable inset with an inhomogeneous background.Using a CT simulation containing the basic physics of metal artifact generation – beam hardening, scatter, and quantum and electronic noise – we generate simulated projection data for the phantom. Here we use a basic MAR algorithm, sinogram inpainting, to test our framework. We reconstruct images using filtered backprojection. To assess MAR algorithm performance, we examine the detectability of a signal at a known location by a Channelized Hoteling Observer (CHO). Our experiment is a two-alternative forced choice study. We use ROC analysis to assess the performance of a MAR algorithm, comparing the AUC with and without use of the MAR algorithm.
Results: Preliminary results show that the algorithm alters artifact appearance with the potential to improve lesion detectability. However, the magnitude of improvement differs for the two CHO setups we investigated, with MAR increasing the mean AUC from 0.73 to 0.79 using four difference-of-Gaussian channels, and 0.90 to 0.91 using 40 Gabor channels.
Conclusion: Our preliminary data confirm the importance of objective assessment methods, suggesting that MAR algorithms can alter image appearance dramatically without a correspondingly large improvement in the actual detectability of a signal, as in the case of the conditions investigated here. We have identified a need for further study to (1) determine the most suitable type and number of channels, and (2) increase study efficiency, possibly by using a signal localization rather than detection task.
Contact Email: