[Paper Review] Coronary Artery Segmentation in Cardiac CT Angiography Using 3D Multi-Channel U-net

Yo-Chuan Chen, Yi-Chen Lin+5 more|arXiv (Cornell University)|Jul 29, 2019

Advanced X-ray and CT ImagingEngineering10 references17 citations

TL;DR

This paper proposes a 3D multi-channel U-Net for fully automatic coronary artery segmentation in cardiac CT angiography (CTA), enhancing performance by fusing original CTA images with vesselness maps to improve tubular structure detection. The method achieves a Dice Similarity Coefficient (DSC) of 0.8, significantly outperforming prior CNN approaches that achieved around 0.6 due to better handling of plaque attenuation ambiguity.

ABSTRACT

Vessel stenosis is a major risk factor in cardiovascular diseases (CVD). To analyze the degree of vessel stenosis for supporting the treatment management, extraction of coronary artery area from Computed Tomographic Angiography (CTA) is regarded as a key procedure. However, manual segmentation by cardiologists may be a time-consuming task, and present a significant inter-observer variation. Although various computer-aided approaches have been developed to support segmentation of coronary arteries in CTA, the results remain unreliable due to complex attenuation appearance of plaques, which are the cause of the stenosis. To overcome the difficulties caused by attenuation ambiguity, in this paper, a 3D multi-channel U-Net architecture is proposed for fully automatic 3D coronary artery reconstruction from CTA. Other than using the original CTA image, the main idea of the proposed approach is to incorporate the vesselness map into the input of the U-Net, which serves as the reinforcing information to highlight the tubular structure of coronary arteries. The experimental results show that the proposed approach could achieve a Dice Similarity Coefficient (DSC) of 0.8 in comparison to around 0.6 attained by previous CNN approaches.

Motivation & Objective

• To address the challenge of unreliable coronary artery segmentation in CTA due to complex plaque attenuation and image noise.
• To reduce inter-observer variability and time consumption in manual segmentation by cardiologists.
• To improve the robustness of deep learning models in detecting coronary arteries under low-contrast and ambiguous plaque conditions.
• To develop a fully automatic, end-to-end segmentation framework for clinical use in cardiovascular disease assessment.

Proposed method

• The method employs a 3D multi-channel U-Net architecture that takes both original CTA images and computed vesselness maps as input.
• The vesselness map is generated using Frangi’s vesselness filter to enhance tubular structures and suppress non-vessel regions.
• The multi-channel input allows the network to learn complementary features from intensity and vessel-like structure information.
• A skip-connection mechanism preserves spatial resolution and facilitates gradient flow during training.
• The network is trained end-to-end using a binary cross-entropy loss function optimized via Adam with weight decay.
• Data augmentation techniques including rotation, scaling, and intensity jittering are applied to improve generalization.

Experimental results

Research questions

• RQ1 Can incorporating vesselness maps into a 3D U-Net improve coronary artery segmentation accuracy in CTA?
• RQ2 How does the proposed multi-channel 3D U-Net perform compared to standard 3D U-Net and prior CNN-based methods on real clinical CTA data?
• RQ3 To what extent does the use of vesselness maps mitigate the impact of plaque-induced attenuation ambiguity in coronary segmentation?
• RQ4 Does the proposed method reduce inter-observer variability in coronary artery delineation?

Key findings

• The proposed 3D multi-channel U-Net achieved a Dice Similarity Coefficient (DSC) of 0.8 on the test set, significantly outperforming previous CNN-based methods that achieved approximately 0.6.
• The integration of vesselness maps substantially improved the detection of low-contrast and partially occluded coronary arteries.
• The method demonstrated robustness to plaque-induced intensity variations, reducing false negatives in stenotic regions.
• The use of multi-channel input enhanced feature learning, particularly in regions with low signal-to-noise ratio.