Purpose: Using an anthropomorphic phantom, quantitatively study noise texture of CT images reconstructed with deep-learning image reconstruction (DLIR, True Fidelity, GE) and compare them with filtered back-projection (FBP) and iterative reconstruction (ASiR-V),
Methods: Axial CT scans of abdomen region of an anthropomorphic phantom (Kyoto Kagaku CTU-41) were acquired from a GE Revolution scanner with four CTDIvol levels ranging from 2.5 mGy (25% of TCM determined mAs) to 46.94 mGy (scanner max mAs). CT slices were reconstructed using all three algorithms, 1.25 mm and 5 mm slice thicknesses, five strength levels of ASiR-V (0% for FBP, 10%, 40%, 80%, and 100%) and three levels of DLIR (low, medium, and high). The 2D noise power spectrum (NPS) was estimated through Fourier analysis of difference images by subtracting two matching images from identical scans. Visual comparison of noise texture was performed between recon algorithms and dose levels via 1D NPS plots. Noise variance (0D NPS) as a quantitative metric was used to study the dependence of noise on algorithm, dose level, and slice thickness.
Results: Image variance had a strong (R^2>0.99) power relationship with dose level for all algorithms, with the power around -0.65. Image variance was reduced by 66% from 1.25 mm to 5 mm slice thickness, by 8% for each 10% of ASiR-V strength increment and by 24% for each DLIR level increment from FBP. The shape of NPS from ASiR-V gradually biased towards lower spatial frequency and deviated away from that of FBP. Our results show a deterioration in spatial resolution with increasing ASiR-V strength. NPS from DLIR had very close shapes to that of FBP, for all three strength levels.
Conclusion: Noise texture with DLIR reconstruction is closer to FBP than ASiR-V. DLIR, as compared to ASiR-V, preserves spatial resolution while reducing imaging noise.
Not Applicable / None Entered.