Discriminability of Non-Gaussian Noise Properties in CT and the Limitation of the Noise Power Spectrum to Describe Noise Texture
K Boedeker1*, D Shin2, L Oostveen3, I Sechopoulos4, C Abbey5, (1) Canon Medical Systems Corporation, Otawara, Japan (2) (1) Canon Medical Systems Corporation, Otawara, Japan (3) Radboud University Nijmegen Medical Centre, Nijmegen, ,NL, (4) Radboud University Medical Centre, Nijmegen, ,(5) University of California - Santa Barbara, Santa Barbara, CA
Presentations
SU-F-201-7 (Sunday, 7/10/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Room 201
Purpose: To demonstrate the impact of higher order, non-Gaussian noise properties, not captured by the Noise Power Spectrum (NPS), on perceived noise texture via a human observer study evaluating abdominal CT reconstruction algorithms.
Methods: A 32 cm water phantom was scanned at two tube current levels on an Aquilion ONE Genesis CT system (Canon Medical Systems, Otawara, Japan). The raw data was reconstructed with three abdominal reconstruction algorithms: a model-based iterative reconstruction (MBIR) algorithm (FIRST), a deep-learning reconstruction (DLR) algorithm (AiCE), and, for one tube current, filtered backprojection (FBP) as a control. Regions of interest (ROIs) (100x100 pixels) were extracted from the image data, and 4th order statistics of each ROI were assessed via excess kurtosis measurement. Pure Gaussian noise counterpart image datasets with the same mean, standard deviation (SD), and NPS as each acquired data condition was also generated by convolving random white noise with the root-NPS of the acquired data. Using a two-alternative forced choice experiment, nine naïve observers were tasked with distinguishing the acquired noise image from its pure Gaussian counterpart.
Results: Excess kurtosis in the image ROIs was 0.01 for FBP, 0.74-0.85 for FIRST, and -0.13-0.21 for AiCE. FBP images appear indistinguishable from their pure Gaussian counterparts (AUC=.54), while MBIR images were readily distinguishable from Gaussian ones (AUC=.98-1). DLR images are more difficult to distinguish from their pure Gaussian counterparts (AUC 0.56-.85), which indicates that it is more similar in perceived texture to Gaussian noise.
Conclusion: This work demonstrates that the appearance of CT noise texture may be dependent on higher orders statistics not captured by the NPS; noise textures with identical NPS and SD can be distinguished based on non-Gaussian properties. Deep-learning reconstruction has lower levels of excess kurtosis and is less readily discriminable from Gaussian noise, despite having a similar NPS to MBIR.
Funding Support, Disclosures, and Conflict of Interest: K Boedeker and D Shin are employees of Canon Medical Systems, Inc C Abbey is a consultant to Canon Medical and an advisor to Izotropic (Breast CT). I Sechopoulos Research agreements: Canon Medical Systems, Siemens Healthcare, ScreenPoint Medical, Sectra Benelux, Volpara Healthcare, Lunit
Keywords
Noise Power Spectrum, Reconstruction, Observer Performance
Taxonomy
IM- CT: Quality Control and Image Quality Assessment
Contact Email