Cycle Consistent Adversarial Denoising Network for Multiphase Coronary CT Angiography

Clinical motivation

• Multiphase acquisition remains important because different cardiac phases may better reveal disease, valve abnormalities, or myocardial motion.The paper notes that cardiac motion information can be essential for evaluating cardiac function.Source: §I · p. 3
• Multiphase tube-current modulation reduces total radiation dose but introduces noise and different contrast levels across cardiac phases.Clinical protocols obtain at least one routine-dose phase that can help radiologists interpret low-dose phases.Source: §I · p. 3

Existing approaches

• Conventional MBIR methods address low-dose noise but require relatively long reconstruction times because of iterative forward and back projections.Deep learning methods learn image statistics directly from data and can enable real-time reconstruction after training.Source: §I · p. 4

Limitations of existing training data

• Supervised denoising is difficult because matched low- and routine-dose images are hard to obtain without additional radiation or motion-corrupted acquisitions.Consequently, many studies use vendor-provided simulated low-dose data, whose noise patterns may differ from real acquisitions and bias training.Source: §I · p. 4

Existing approaches

• GAN-based denoising can match routine-dose image distributions, but mode collapse may generate features absent from the input and fail to preserve meaningful pairing.GAN mode collapse can produce limited or identical outputs regardless of the input.Source: §I · p. 4

Proposed approach

• The proposed network trains inverse mappings between low- and routine-dose domains using cyclic consistency and identity loss.The inverse path favors one-to-one correspondence, while identity loss constrains outputs in the target domain.Source: §I · p. 5

Reported outcome

• Experimental results indicate robustness to cardiac motion and contrast changes without creating artificial features.The network generates denoised samples from noisy inputs rather than random noise, reducing the likelihood of mode collapse.Source: §I · p. 5

• Low-dose CT is represented as domain A, while routine-dose CT is represented as domain B with distributions P_A and P_B.Source: §II · pp. 5–6
• Generators G_AB and G_BA map images between the low-dose and routine-dose domains in opposite directions.Source: §II · p. 5
• The framework contains two generators, G_AB and G_BA, and two discriminators, D_A and D_B.Source: §II · p. 6
• Discriminators D_A and D_B distinguish measured input images from synthesized images produced by the generators.Source: §II · p. 6
• Generators and discriminators are trained simultaneously by optimizing an overall loss combining adversarial, cyclic, and identity terms.Source: §II · p. 6

II.A.1. Adversarial loss

• G_AB and D_B are trained through a min-max adversarial objective derived from the original GAN formulation.Source: II.A.1. Adversarial loss · p. 7
• The adversarial loss trains G_AB to make denoised low-dose images resemble routine-dose images while D_B distinguishes generated from routine-dose images.Source: II.A.1. Adversarial loss · p. 7
• The original adversarial loss was unstable during training, so its log-likelihood function was replaced with a least-squares loss.Source: II.A.1. Adversarial loss · p. 7
• The revised formulation expresses training as two minimization problems based on the least-squares GAN objective.Source: II.A.1. Adversarial loss · p. 7
• The adversarial loss guides the generator toward denoised images classified as real routine-dose images.Source: II.A.1. Adversarial loss · p. 7
• The discriminator simultaneously learns to distinguish denoised images from routine-dose images.Source: II.A.1. Adversarial loss · p. 7
• A corresponding adversarial loss is applied to G_BA, which generates noisy CT images in the reverse domain direction.Source: II.A.1. Adversarial loss · p. 7

II.A.2. Cyclic loss

• Adversarial losses produce realistic denoised and noisy images but do not guarantee that the two generators form an inverse relation.Source: II.A.2. Cyclic loss · p. 7
• The cyclic loss passes an image through both generators and penalizes deviation from the original input using the l1-norm.Source: II.A.2. Cyclic loss · pp. 7–8
• The cyclic constraint encourages G_AB and G_BA to act as approximate inverses, including G_BA(G_AB(x_A)) ≈ x_A.Source: II.A.2. Cyclic loss · p. 8
• Multiphase CTA can contain dose-phase mismatches because arrhythmia changes the expected relationship between cardiac phase and dose modulation.Source: II.A.3. Identity loss · p. 8
• A systolic-phase image expected to be low dose may instead have full-dose noise characteristics after cardiac-cycle mismatch.Source: II.A.3. Identity loss · p. 8

II.A.3. Identity loss

• Identity loss trains G_AB and G_BA not to change inputs that are already clean or already at the low-dose level.Source: II.A.3. Identity loss · p. 8
• Identity constraints make correctly generated target-domain images fixed points of the generator.Source: II.A.3. Identity loss · p. 8
• The identity constraint is reported as important for avoiding artificial features.Source: II.A.3. Identity loss · p. 8

• The two generator architectures are illustrated in Fig. 3 and are optimized for low-dose CT image denoising.Source: II.B. Network architecture · p. 9
• The generators process image inputs directly rather than prior-work wavelet-transform coefficients.Source: II.B. Network architecture · p. 9
• Each generator begins with 128 3×3 convolution kernels and uses six modules containing convolution, batch-normalization, ReLU, and bypass layers.Source: II.B. Network architecture · p. 9
• A concatenation layer combines module inputs with the last-module output to provide multiple signal representations and gradient paths.Source: II.B. Network architecture · p. 9
• The final convolution layer uses 15 sets of 3×3×128 kernels, followed by an end-to-end bypass connection for estimating the noise-free image.Source: II.B. Network architecture · p. 9
• The discriminators use a PatchGAN architecture with a 70×70 receptive field to classify image patches as real or synthesized.Source: II.B. Network architecture · pp. 9–10
• The discriminator has five convolution layers and aggregates l2-loss over 56×56 patches, supporting arbitrarily sized images.Source: II.B. Network architecture · p. 10
• The study cohort included 50 CT scans of mitral valve prolapse patients and 50 CT scans of coronary artery disease patients.Source: III.A. Data: Cardiac CT scans · p. 10

Optimization

• Training minimized loss function (2) with λ = 10 and γ = 5 using ADAM optimization with β1 = 0.5 and β2 = 0.999.Source: III.B. Training details · p. 11

Training schedule

• The network was trained for 160 epochs, using learning rate 0.0002 for the first 100 epochs before linear decay to zero.Early stopping was performed at 160 epochs as a regularization measure.Source: III.B. Training details · p. 11

Training configuration

• Training used 56 × 56 patches and mini-batches of 10, with randomly Gaussian-initialized kernels and generator–discriminator updates at each iteration.Source: III.B. Training details · p. 11

Preprocessing

• Input low-dose and target routine-dose CT images were intensity-normalized, with inputs scaled to the range [−1, 1].Source: III.B. Training details · p. 11

Quantitative analysis

• The evaluation measured image noise and SNR within four anatomical structures using structure-confined regions of interest.The structures were the ascending aorta, left ventricular cavity, left ventricular septum, and proximal right coronary artery.Source: III.C.2. Quantitative analysis · p. 12

Restoration quality

• The proposed method reduced noise while preserving texture and edge information in low-dose cardiac CT images.Source: IV.A. Qualitative evaluation · p. 14

Robustness and artifact control

• The method did not create artificial features and was robust across mitral valve prolapse and coronary artery disease cases.Source: IV.A. Qualitative evaluation · p. 14

Identity behavior

• The generator GAB acted as the identity for target-domain images, including cases with noise levels similar to the target images.Source: IV.A. Qualitative evaluation · pp. 15–16

Comparison with ADMIRE

• Compared with ADMIRE, both methods reduced noise, but ADMIRE showed edge loss, over-smoothing, and remaining streaking artifacts that were not observed with the proposed method.Source: IV.A. Qualitative evaluation · pp. 17–18

Visual grading

• All visual scores were significantly higher for denoising CT than original CT images (P < 0.001).Source: IV.B. Visual grading score and SNR… · p. 18

Noise and SNR

• Image noise decreased and SNR significantly increased in denoising CT images (P < 0.05).Source: IV.B. Visual grading score and SNR… · pp. 18–19
• SNR changes in the left ventricular cavity were not statistically significant (P = 0.96).The left ventricular cavity used the largest region of interest and contained contrast-enhanced blood pool.Source: IV.B. Visual grading score and SNR… · p. 19

Application to AAPM Data Set

• The AAPM dataset contained abdominal CT images from ten patients, with eight patients used for training and validation and two for testing.Source: IV.C. Application to AAPM Data Set · p. 19
• Training on AAPM data was unsupervised, with input and target images randomly selected from the entire dataset.Source: IV.C. Application to AAPM Data Set · p. 19

Comparison with supervised learning

• The proposed unsupervised method produced better images than the supervised learning approach in Fig. 10(b), although some artifacts remained in other results.Source: IV.C. Application to AAPM Data Set · p. 19

Ablation analysis

• All reconstruction results reduced noise and maintained edge information, whereas other methods generated structures absent from the input images.Source: IV.D. Ablation study · p. 19
• Excluding identity loss and cyclic loss caused heart-shape deformation and removal of structures compared with the proposed method.Source: IV.D. Ablation study · p. 20
• The experiments supported the importance of cyclic and identity losses because artificial features can confuse radiologists during diagnosis.Source: IV.D. Ablation study · p. 20

Training stability

• LSGAN and cycle loss produced stable convergence during training, with LGAB and LDB approaching 0.25.The cyclic loss decreased steadily and converged, which the authors interpreted as evidence of well-done network training.Source: §V · pp. 21–22

GAN limitations and safeguards

• Cycle consistency and identity loss reduce mode collapse and prevent artificial features from appearing in outputs.The authors report no cases where outputs from distinct inputs were identical.Source: §V · p. 22

Limitations and generalizability

• The study mainly evaluates multiphase CTA using data reconstructed with the same B26f cardiac kernel.Generalization to different kernels or organs remains unclear, and broader validation is required.Source: §V · p. 22

Clinical practicality

• Extensive clinical evaluation is still required before major acquisition-protocol changes, whereas the proposed method retains current protocols.The authors describe this compatibility as more readily acceptable in the current clinical setting.Source: §V · p. 22
• The approach enhances images within current acquisition protocols and does not require exactly matched low- and routine-dose image pairs.The authors position enhanced images as additional information for radiologists and note that matched clinical pairs are difficult to obtain.Source: §V · p. 22

VI. CONCLUSION

• The proposed network denoises low-dose coronary CT images while maintaining texture and edge information.Experimental results confirmed this behavior for input low-dose CT images.Source: §VI · p. 23
• Unlike supervised approaches, the network does not require exactly matched low- and routine-dose images.It learns image distributions from high-dose cardiac phases instead.Source: §VI · p. 23
• Cyclic consistency and identity loss are used to prevent artificial features that are absent from the input image.The conclusion contrasts this design with GAN-based networks prone to generating artificial features.Source: §VI · p. 23
• Routine-dose inputs remain unchanged, supporting that the algorithm correctly learned the noise.This behavior was reported when routine-dose images were used as inputs.Source: §VI · p. 23
• Visual grading analysis confirmed a significant increase in diagnostic image quality.The authors also describe the method as effective and practicable.Source: §VI · p. 23
• The method may be applicable to other CT acquisition protocols with dynamic tube current.This is presented as a potential broader application of the proposed approach.Source: §VI · p. 23