Ability of Deep Learning Image Reconstruction to Preserve Detail and Remove Noise in Coronary Computed Tomography Angiography: A Clinical Analysis

PURPOSE: To analyze image quality over 3 levels (low, medium, and high) of deep learning image reconstruction (DLIR), evaluate whether edge detail is lost with higher denoising levels. MATERIALS AND METHODS: Nineteen subjects scanned with the institute's clinical coronary CT angiography (CCTA) protocols were retrospectively reconstructed with low, medium, and high levels of DLIR. Two readers assessed image quality on a 5-point Likert scale for structural confidence in relevant structures, low contrast detectability, noise, spatial resolution, artifacts, and overall diagnostic quality. The noise texture of images was also assessed using a non-Likert 5-point scale of "plastic" to "sandy." Reader agreement was assessed with pairwise linearly weighted Cohen Kappa. Objective analysis was performed with contrast-to-noise (CNR) measurements of the aorta, left main artery, and right coronary artery. Friedman and Dunnett tests were used to assess statistical differences in image quality and CNR. To assess detail loss, edge detail preservation was assessed with different images between DLIR levels. RESULTS: Observations of the difference images between DLIR images and the histograms of edge and nonedge pixels demonstrated that negligible edge detail was lost between the DLIR reconstructions. High DLIR CNR was significantly (P≤0.006) higher than low and medium DLIR CNR. Medium DLIR CNR was significantly (P≤0.02) higher than low DLIR CNR. Reader agreement was moderate (κ=0.42). High DLIR scores were significantly superior (P≤0.002) to low DLIR scores for most structural confidence structures, low contrast detectability, noise, spatial resolution, artifacts, and overall diagnostic quality, and significantly superior (P≤0.03) to medium DLIR scores for all categories except spatial resolution (P=0.66). CONCLUSION: By both objective and subjective measures, the higher levels of DLIR resulted in improved image quality without loss of edge details.