Cycle Consistent Adversarial Denoising Network for Multiphase Coronary CT Angiography (1806.09748v3)

Abstract: In coronary CT angiography, a series of CT images are taken at different levels of radiation dose during the examination. Although this reduces the total radiation dose, the image quality during the low-dose phases is significantly degraded. To address this problem, here we propose a novel semi-supervised learning technique that can remove the noises of the CT images obtained in the low-dose phases by learning from the CT images in the routine dose phases. Although a supervised learning approach is not possible due to the differences in the underlying heart structure in two phases, the images in the two phases are closely related so that we propose a cycle-consistent adversarial denoising network to learn the non-degenerate mapping between the low and high dose cardiac phases. Experimental results showed that the proposed method effectively reduces the noise in the low-dose CT image while the preserving detailed texture and edge information. Moreover, thanks to the cyclic consistency and identity loss, the proposed network does not create any artificial features that are not present in the input images. Visual grading and quality evaluation also confirm that the proposed method provides significant improvement in diagnostic quality.

• The paper introduces an unsupervised CycleGAN-based denoising network to map low-dose coronary CTA images to routine-dose quality.
• It employs adversarial, cyclic, and identity loss functions to preserve texture fidelity and ensure authentic image transitions without paired datasets.
• Evaluations demonstrate improved signal-to-noise ratios and diagnostic clarity compared to traditional iterative reconstruction methods.

Cycle Consistent Adversarial Denoising Network for Multiphase Coronary CT Angiography

The paper presents a methodology for addressing the noise degradation challenges associated with multiphase coronary CT angiography (CTA), specifically in the context of low-dose image acquisition. In multiphase CTA, images are acquired with varying radiation doses to mitigate the overall exposure but inevitably result in quality loss during low-dose phases. The novel approach introduced combines unsupervised deep learning with cycle-consistent adversarial networks (CycleGAN) to effectively denoise images.

Methodology

The authors propose an unsupervised learning framework that leverages the cycle-consistent adversarial denoising network (CADN). This network aims to learn the mapping from low-dose to routine-dose cardiac phases by considering the intrinsic relationships between the two. Unlike traditional supervised techniques requiring paired datasets, this method circumvents the challenge by utilizing the cycle consistency property which allows learning even with unpaired data. The network involves generators for transitioning between phases and discriminators evaluating the authenticity of these transitions.

The primary components of the loss function for training the network include:

1. Adversarial Loss: Utilizes an adversarial discriminative approach that refines the generator to produce outputs that resemble the target dose images closely and discriminators to distinguish between authentic and synthesized images.
2. Cyclic Loss: Enforces the generators in the network to be inverses of one another, maintaining consistency and reducing the risk of mode collapse, which can lead to artificial generation of features unaligned with the input data.
3. Identity Loss: Ensures that when images at routine doses are input to the generator designed for low-dose inputs, they remain unchanged. This property is crucial for maintaining the authenticity of high-quality images.

The generators and discriminators are implemented with specific architectural considerations derived from previous work on optimizing low-dose CT denoising tasks.

Results and Evaluation

Quantitative and qualitative assessments reveal considerable improvements in denoised images compared to conventional model-based iterative reconstruction methods, such as the ADMIRE algorithm. The cycle consistent adversarial framework showcases its efficacy across a validated dataset of patients with different cardiac conditions, maintaining texture fidelity and edge definition in denoised images along with visual grading analysis.

Furthermore, the framework demonstrated adaptability in preserving high-quality image characteristics where applicable and showcased robustness against noise induction without fabricating non-existent structures. Statistical measures confirmed improved diagnostic quality with significant gains in signal-to-noise ratios and reduced image noise statistics.

Implications and Future Directions

The implications of this research are substantial for clinical radiography, especially for dynamic tube current modulation during CT protocols in cardiology. By efficiently using existing acquisition protocols and augmenting the diagnostic quality without the necessity for matched datasets, this approach promises to preserve patient quality care while mitigating radiation exposures.

Future work could explore expanding the methodology to different anatomical domains, alternative reconstruction kernels, and investigating its generalization for distinct clinical tasks beyond coronary CTA. The evolution of unsupervised learning frameworks like cycle GAN presents a promising trajectory for advancing non-invasive diagnostic imaging with minimal reliance on augmented data sets and direct supervised methods. The potential for scalability and adaptability in diverse medical imaging scenarios renders this work of significant long-term value in healthcare technology and product development.