Deep learning-based reconstruction improves image quality in low-dose head CT angiography.

Abstract

Objective: This study aimed to compare the image quality of filtered back projection (FBP), adaptive statistical iterative reconstruction-Veo (ASIR-V) and the deep learning image reconstruction (DLIR) algorithms in low-dose head CT angiography (CTA).

Methods: This prospective study was conducted on 25 patients undergoing head CTA using a 256-slice CT scanner. Patients received 25 mL of iodine contrast (Iopromide, 370 mg I/mL, 3.0 mL/s). Images were reconstructed using DLIR with high settings (DLIR-H) and medium settings (DLIR-M), FBP, and ASIR-V with a blending factor of 50% (ASIR-V 50%). CT values, standard deviations, signal-to-noise ratios (SNR), and contrast-to-noise ratios (CNR) were measured at the basal ganglia, posterior cranial fossa, center of semiovale, and middle cerebral artery. The edge rise slope (ERS) of the middle cerebral artery rim was measured to assess vessel clarity. Image noise, vessel edge definition, and overall quality were scored on a 5-point scale, while sharpness and clarity were rated on a 4-point scale.

Results: FBP images exhibited the highest image noise, as reflected by SD values. DLIR, especially DLIR-H, showed superior noise reduction compared to ASIR-V 50%. SNR followed this trend: FBP < ASIR-V 50% < DLIR-M < DLIR-H. Spatial resolution, measured by ERS for vessel wall clarity, was higher in DLIR images compared to in ASIR-V 50%. DLIR outperformed conventional iterative algorithms in balancing noise reduction and edge clarity, with both DLIR-M and DLIR-H achieving better subjective scores for noise, edge definition, and sharpness than ASIR-V 50% and FBP.

Conclusion: DLIR in low-dose head CTA could reduces image noise, preserve natural texture, and enhance image clarity compared with ASIR-V and FBP methods.