Purpose: To compare radiologist preference for a model-based iterative reconstruction algorithm (FIRST) to a statistical-based iterative reconstruction algorithm (AIDR 3D) using clinical images.
Methods: IRB approval was obtained to retrospectively collect and post-process 60 clinical CT exams from lung screening, pulmonary embolism, and non-contrast abdomen protocols (20 from each protocol). These protocols represented three image-quality challenges: low dose, high spatial resolution, and low contrast imaging, respectively. The cases were reconstructed with AIDR 3D and FIRST using default processing settings. Each image pair was presented side-by-side on a diagnostic workstation to three radiologists, with the position of imaging set on left or right randomized. Five image-quality-related categories and four exam-specific clinical categories were rated, with the radiologists indicating a preference strength (strong, slight, or none) and direction (left or right). A two-sided test Wilcoxon signed-rank test was used to analyze the scores in each category, with a p-value < 0.05 considered a significant preference. Analysis was performed for each category and radiologist individually, as well as combined across radiologists.
Results: FIRST was not preferred in any category when radiologist scores were combined, and only rarely when separated. For the “overall preference” category, AIDR 3D was preferred for lung screening by two out of three radiologists; the third had no preference. There was no preference between algorithms for pulmonary embolism. AIDR 3D was preferred by one radiologist for non-contrast abdomen, while the two remaining had no preference. FIRST increased breast shield artifacts, but reduced metal implant artifacts.
Conclusion: The radiologist image quality study indicated preference for AIDR 3D over FIRST for two out of three imaging quality challenges, with no preference for the third. It is likely that noise texture influenced radiologist preference. Careful clinical implementation of reconstruction algorithm is necessary as FIRST may not be preferred for all clinical applications.
Funding Support, Disclosures, and Conflict of Interest: This project was supported in part by Canon Medical Systems USA, Inc. The content of this work is solely the responsibility of the authors and does not represent the official views of Canon Medical Systems. All other authors have nothing to disclose.