Exhibit Hall | Forum 8
Purpose: Two vendor-specific deep learning iterative reconstruction algorithms (DLIR) for CT were assessed for image quality impact towards dose reduction feasibility.
Methods: An ACR CT phantom was imaged in decrementing fixed mA levels from 300 mAs to 10 mAs in steps of 20 mAs. The phantom was first imaged using a Toshiba Aquilion One Genesis scanner (Canon Medical), and images were reconstructed using AiCE DLIR. The phantom was then imaged using a Revolution Apex scanner (GE Healthcare), and images were reconstructed using TrueFidelity DLIR. Noise power spectra (NPS), CNR, and modulation transfer function (MTF) were calculated for each phantom acquisition. ANOVA statistics were calculated to assess noise magnitude and frequency differences.
Results: The mean frequency (f_avg) of each NPS spectra were plotted and AiCE demonstrated a logarithmic reduction in noise frequency as a function of mAs [f_avg = 0.08*ln(mAs)+0.14; R2=0.957]; whereas, TrueFidelity demonstrated no meaningful regression with a nearly uniform f_avg response (0.37±0.01 lp/cm) as a function of mAs. For AiCE, a statistically significant difference was calculated for f_avg values below 90 mAs compared to those above 90 mAs (p = 0.0089); no such difference was calculated for TrueFidelity. CNR was greater at all mAs levels for AiCE. The calculated frequency at the 50% MTF was assessed at all mAs levels for TrueFidelity resulting in no regression (0.47±0.01 lp/cm), and for AiCE a log regression resulted [f=0.1*ln(mAs)+0.22; R2=0.980].
Conclusion: Two DLIR algorithms were assessed for image quality as a function of decreasing mAs. AiCE demonstrated a strong reduction in noise magnitude and a logarithmic reduction in noise frequency (i.e., texture) content; whereas, TrueFidelity demonstrated less noise reduction without a change in noise frequency (i.e., texture) content. MTF was shown to improve logarithmically at all levels of mAs for AiCE whereas MTF was constant for TrueFidelity.