Influence of ClearInfinity Algorithm Weight on the Quality of Virtual Monoenergetic Images in Cranial Spectral Computed Tomography Angiography

Objective: To investigate how the weights of the deep learning–based ClearInfinity (CI) algorithm affect the quality of 55keV virtual monoenergetic images (VMIs) in cranial spectral computed tomography angiography (CTA), and to provide a basis for clinical optimization of reconstruction parameters. Methods: A total of 38 patients who had undergone cranial spectral CTA were retrospectively enrolled. The original data were reconstructed using the CI algorithm with weights of 10%, 30%, 50%, 70%, and 90% to obtain five groups of 55keV VMIs. Objective evaluation was performed by measuring the vascular CT values, gray matter CT values, background noise (BN), signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR). Double-blind subjective scoring of image noise, vascular edges, and detailed display was performed by two radiologists. Results: No statistically significant differences were observed in the CT values of blood vessels (right internal carotid artery, right middle cerebral artery, and basilar artery) or gray matter among the different weights (F=0.787, 0.525, 0.650, and 2.979, respectively; all P>0.05). However, pairwise comparisons showed statistically significant differences in BN, SNR, and CNR among the weight groups (F=736.676, 608.871, and 580.777, respectively; all P<0.05). BN gradually decreased, whereas SNR and CNR gradually increased with increase in algorithm weight. Subjective scores of the five groups showed statistically significant differences (\chi^2 =77.135, P<0.05), with the average scores from high to low being 4.95±0.22 (50% CI), 4.87±0.36 (70% CI), 4.74±0.48 (90% CI), 4.66±0.58 (30% CI), and 4.03±0.62 (10% CI). Conclusion: The CI algorithm weight influences image quality: relatively low weights compromise image quality owing to excessive noise, whereas relatively high weights cause over-smoothing and blurring of the image, resulting in the loss of fine structures. Therefore, this study recommends a 50% CI algorithm weight for reconstructing 55 keV VMIs in cranial spectral CTA.