Deep Learning Denoising Algorithm for Improved Assessment of Coronary Arteries in Transcatheter Aortic Valve Implantation CT Imaging.
Authors
Affiliations (7)
Affiliations (7)
- Diagnostic and Interventional Radiology Unit, BIOMORF Department, University Hospital Messina, 98124 Messina, Italy (L.R.M.L., T.D., S.M.).
- Department of Neuroradiology, University Medical Center Mainz, Johannes Gutenberg University, 55131 Mainz, Germany (S.S., M.A.B., A.E.O.).
- Diagnostic and Interventional Radiology Unit, BIOMORF Department, University Hospital Messina, 98124 Messina, Italy (L.R.M.L., T.D., S.M.). Electronic address: [email protected].
- Division of Experimental Imaging, Department of Diagnostic and Interventional Radiology, University Hospital Frankfurt, 60590 Frankfurt, Germany (I.Y., V.K., L.D.G., S.S.M., J-E.S., K.E., L.S.A., C.B.); Department of Diagnostic and Interventional Radiology, University Hospital Frankfurt, 60590 Frankfurt, Germany (I.Y., V.K., L.D.G., S.S.M., J-E.S., K.E., L.S.A., C.B.).
- Department of Diagnostic and Interventional Radiology, University Hospital Frankfurt, 60590 Frankfurt, Germany (I.Y., V.K., L.D.G., S.S.M., J-E.S., K.E., L.S.A., C.B.).
- Research Institute, ClariPi, Seoul 03086, Republic of Korea (C.A., J.H.K.).
- Research Institute, ClariPi, Seoul 03086, Republic of Korea (C.A., J.H.K.); Department of Radiology, Seoul National University Hospital, 101 Daehakro, Jongnogu, Seoul 03080, Republic of Korea (J.H.K.); Department of Bioengineering, Graduate School of Convergence Science and Technology, Seoul National University, Seoul 08826, Republic of Korea (J.H.K.).
Abstract
To assess the impact of a deep learning-based noise reduction (DLD) technique on image quality and diagnostic accuracy for the evaluation of coronary arteries in transcatheter aortic valve implantation (TAVI) CT imaging. Two hundred patients with severe aortic stenosis who underwent CT scans for pre-TAVI planning between October 2022 and April 2024 were retrospectively enrolled. Conventional images were reconstructed and denoised images were generated using dedicated software. Objective image quality was evaluated by measuring the mean Hounsfield unit (HU) and standard deviation (SD) in regions of interest within the aortic root, coronary arteries, and subcutaneous fat to calculate signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR). For subjective assessment, two readers used a 5-point Likert scoring system to evaluate sharpness, noise, contrast and overall image quality. The diagnostic performance of both datasets was assessed using invasive coronary angiography as reference standard. Denoised reconstructions showed significantly higher SNR (37.5±12.8 vs.12.3±4.1) and CNR (45.3±15.4 vs. 14.7±4.4), and lower noise (16.9±7.9 vs. 47.9±11.6 HU) (all p<0.001). Subjective assessment demonstrated that denoised images received the highest score for sharpness, noise, contrast and overall image quality (all p<0.001). For the evaluation of diagnostic accuracy, a total of 800 vessels and 1787 segments were analyzed. The per-segment diagnostic performance of the DLD for detection of CAD revealed an AUC of 90% (95% CI: 88.5-91.3), with accuracy of 93.9% (95% CI: 92.7-95), 85.7% (95% CI: 78.7-90.4) sensitivity and 94.7% (95% CI: 93.5-95.7) specificity, in the absence of a statistically significant difference compared with the evaluation performed on standard images (p=0.056). The DLD substantially improves image quality without affecting diagnostic accuracy for the evaluation of coronary arteries in patients undergoing pre-TAVI CT scans.