Image Quality Improvement and Artificial Intelligence Performance in Pulmonary Embolism Detection at Deep Learning Reconstruction-Based Ultra-low Radiation Dose CT Pulmonary Angiography.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, Shanghai Geriatric Medical Center, Shanghai 200000, China (J.L., L.S., Z.B., M.Z., M.W.).
- Department of Radiology, Zhongshan Hospital, Fudan University, Shanghai 200000, China (C.Z., X.Y., M.Z.).
- CT Business Unit, Canon Medical Systems (China), Shanghai 200000, China (Z.Z.).
- Department of Radiology, Shanghai Geriatric Medical Center, Shanghai 200000, China (J.L., L.S., Z.B., M.Z., M.W.); Department of Radiology, Zhongshan Hospital, Fudan University, Shanghai 200000, China (C.Z., X.Y., M.Z.).
- Department of Radiology, Shanghai Geriatric Medical Center, Shanghai 200000, China (J.L., L.S., Z.B., M.Z., M.W.). Electronic address: [email protected].
Abstract
To evaluate the image quality of deep learning reconstruction (DLR)-based ultra-low dose (ULD) CT pulmonary angiography (CTPA) images and determine whether the artificial intelligence (AI) software can improve the diagnostic performance of radiologist for detecting pulmonary embolism (PE) with ULD images. This prospective two-center study enrolled 144 patients with suspected PE who underwent CTPA from July to October 2024. Patients were randomized into two groups equally. Images in the routine-dose (RD) group were reconstructed using hybrid-iterative reconstruction (HIR), while ULD images were reconstructed using HIR and DLR. A subset of 56 participants (1:1 PE to non-PE ratio) in ULD group was randomly selected and evaluated by three radiologists with and without AI software. Reference standard was established by expert consensus. Interrater reliability was determined by intraclass correlation coefficient (ICC). The diagnostic results and interpretation times were recorded. There were no significant differences in demographics between the two groups. ULD-DLR images exhibited significantly higher objective and subjective image quality compared to both RD-HIR and ULD-HIR images. Interobserver agreement was moderate for RD-HIR (ICC=0.77) and excellent for ULD-DLR images (ICC=0.84). For radiologist detection of PE assisted by AI, both ULD-HIR and ULD-DLR cohorts exhibited near-perfect accuracy, outperforming unassisted readings (sensitivity 79.8% vs. 91.7% and specificity 95.5% vs. 99.2% in ULD-HIR; sensitivity 90.5% vs. 96.4% and specificity 95.8% vs. 100.0% in ULD-DLR). AI assistance reduced interpretation time by 19.7% for ULD-HIR and 15.6% for ULD-DLR scans. The effective dose of ULD group was decreased by 74% compared to RD group. DLR can maintain the CTPA image quality even at ultra-low dose level, further ensuring the accuracy and efficiency of AI-assisted PE diagnosis while improving radiation safety.