Room: AAPM ePoster Library
Compare deep learning-based CT reconstruction (AiCE) to FBP, statistical-based (AIDR3D), and model-based (FIRST) iterative reconstruction.
CT images for 19 (3-19 YO) patients, acquired on an Aquilion ONE Genesis CT, were axially-reconstructed at 0.5mm and 3mm thicknesses. A non-prewhitening matched mathematical observer model with eye filter (d’NPWE) was used to characterize SNR of varying size objects (0.5-10mm) at different CT contrast levels (50, 150, 250, and 350HU) above contrast enhanced liver parenchyma (~100HU). To calculate d’NPWE, for each reconstruction algorithm, TTF was calculated using a Catphan 600 phantom, and NPS was calculated by sampling noise characteristics from uniform regions of each patients’ liver. Object detection sensitivity was estimated using AUC calculations. AUC results for FBP, FIRST, and AiCE were normalized to AIDR3D, the clinically employed reconstruction algorithm. Three pediatric radiologists provided a multi-reader multi-case (MRMC) analysis of images from four anatomical locations, and scored image quality. Linear mixed models and Tukey Posthoc analysis studied the effect of reconstruction and the interaction between anatomical location and reconstruction.
NPS magnitude for 3mm AiCE images were an average 58% lower (range: 45-70%) than 3mm AIDR3D images. Noise texture of AiCE agrees to better than 28% with AIDR3D compared to 50% for FIRST. On average, AiCE had greater SNR for all object sizes and contrast levels. AiCE images had greater detection sensitivity. AiCE 0.5mm SNR agreed with 3mm AIDR3D to better than 0.4% demonstrating a potential for ~45-70% dose reduction if image quality was maintained constant as currently implemented with AIDR3D. For each observer in the MRMC analysis, AiCE scored statistically better for image quality, detection sensitivity, and observer preference at all anatomical locations.
AiCE demonstrates substantial object detection improvement with less noisy images, and noise texture comparable with AIDR3D for substantial dose reduction potential.