3-D Neural denoising for low-dose Coronary CT Angiography (CCTA). (December 2018)

• Record Type: Journal Article
• Authors: Green, Michael; Marom, Edith M.; Konen, Eli; Kiryati, Nahum; Mayer, Arnaldo
• Is Part Of: Computerized medical imaging and graphics. Volume 70(2018)
• Page Start: 185
• Page End: 191
• Publication Date: 2018-12
• Subjects: Coronary CT angiography -- Low-Dose CT -- Denoising -- Patches -- Convolutional neural networks; Diagnostic imaging -- Periodicals; Imaging systems in medicine -- Periodicals; Diagnosis, Radioscopic -- Data processing -- Periodicals; Diagnostic Imaging -- Periodicals; Imagerie pour le diagnostic -- Périodiques; Diagnostic imaging; Periodicals; Electronic journals; Electronic journals; 616.0754
• DOI: 10.1016/j.compmedimag.2018.07.004
• Languages: English
• ISSNs: 0895-6111

Abstract

Highlights: Denoising by 3-D neural networks better exploits the 3-D nature of the body. Low dose coronary CT angiography is better denoised by 3-D rather than 2-D networks. 3-D fully convolutional neural networks can denoise full CT scans in a single piece. Abstract: CCTA has become an important tool for coronary arteries assessment in low and medium risk patients. However, it exposes the patient to significant radiation doses, resulting from high image quality requirements and acquisitions at multiple cardiac phases. For widespread use of CCTA for coronary assessment, significant reduction of radiation exposure with minimal image quality loss is still needed. A neural denoising scheme, relying on a fully convolutional neural network (FCNN) architecture, is developed and applied to noisy CCTA. In contrast to previously published methods, the proposed FCNN is trained directly on 3-D CT data patches (blocks), implementing 3-D convolutions. Considering that anatomy is inherently tridimensional, the proposed 3-D approach may better capture and enforce inter-slice continuity of tiny structures. While training is performed on individual blocks, whole input scans can be fed and denoised in one piece, thus leveraging the fully convolutional architecture to maximize processing speed. The proposed method is compared to state-of-the-art denoising algorithms on a dataset of 45 CCTA scans. Low-dose scans are simulated by synthetic Poisson noise applied to the sinogram corresponding to a 90% reduction in radiation dose. The average feature similarity score (0.864) and the peak signal-to-noise ratio (41.47) obtained for the proposed algorithm outperformed the compared methods while requiring significantly shorter processing time. A set of 2-D FCNNs, structurally similar to the proposed 3-D network, are also implemented to demonstrate contribution of the additional dimension to the improved denoising. For further validation of the method coronary reconstruction using the Intellispace cardiac tool (Philips, Holland) is performed both on a real noisy CCTA scan and on the denoised scan using the proposed method. It is shown that the cardiac tool succeeds in reconstructing more coronaries using the scan denoised by the proposed method. The obtained results suggest the proposed method provides an efficient and powerful approach to low-dose CCTA denoising.