Adversarial and Score-Based CT Denoising: CycleGAN vs Noise2Score

TL;DR

This work compares two top-tier, pair-free approaches for CT denoising: a CycleGAN-based residual translator and a Noise2Score denoiser. The CycleGAN employs a standard U-Net backbone with a residual learning path and uses adversarial, cycle-consistency, and identity losses to translate LDCT to NDCT-like images, achieving the highest absolute PSNR/SSIM and unseen-set scores under a shared protocol. Noise2Score relies on Amortized Residual Denoising Autoencoder (AR-DAE) to estimate the score with Tweedie inference at test time, delivering substantial gains from very noisy inputs despite not reaching CycleGAN's absolute metrics. The results suggest a practical guideline: use CycleGAN when unpaired domain transfer is feasible to maximize final image quality, and use Noise2Score when clean training targets are unavailable to obtain robust, pair-free denoising with strong relative improvements.

Abstract

We study CT image denoising in the unpaired and self-supervised regimes by evaluating two strong, training-data-efficient paradigms: a CycleGAN-based residual translator and a Noise2Score (N2S) score-matching denoiser. Under a common evaluation protocol, a configuration sweep identifies a simple standard U-Net backbone within CycleGAN (lambda_cycle = 30, lambda_iden = 2, ngf = ndf = 64) as the most reliable setting; we then train it to convergence with a longer schedule. The selected CycleGAN improves the noisy input from 34.66 dB / 0.9234 SSIM to 38.913 dB / 0.971 SSIM and attains an estimated score of 1.9441 and an unseen-set (Kaggle leaderboard) score of 1.9343. Noise2Score, while slightly behind in absolute PSNR / SSIM, achieves large gains over very noisy inputs, highlighting its utility when clean pairs are unavailable. Overall, CycleGAN offers the strongest final image quality, whereas Noise2Score provides a robust pair-free alternative with competitive performance. Source code is available at https://github.com/hanifsyarubany/CT-Scan-Image-Denoising-using-CycleGAN-and-Noise2Score.

Figures (8)

• Figure 1: Residual-learning overview for the CycleGAN generator: the network predicts a noise residual that is added to the input via a global skip to produce the denoised image.
• Figure 2: Comparison of U-Net generator variants used within the CycleGAN framework for CT denoising. (Top) The standard U-Net architecture augmented with a global residual path, where the network learns a residual noise component added back to the input image to produce the denoised result. (Bottom) The Attention U-Net variant, which incorporates attention gates (AG) along skip connections to enhance feature selection and focus on anatomically relevant regions. In our experiments, additional U-Net variants were also explored, including a Residual Dense U-Net (ResU-Net) that introduces local residual connections within each convolutional block.
• Figure 3: PatchGAN discriminator architecture with progressive $4\times4$ convolutions. Each convolutional layer doubles the feature depth and reduces spatial resolution, enabling the network to classify local image patches rather than the entire image.
• Figure 4: Comprehensive overview of the loss functions used in the proposed CycleGAN-based residual learning framework for CT denoising. The adversarial losses ($\mathcal{L}_{adv,F}$ and $\mathcal{L}_{adv,Q}$) encourage each generator to produce realistic images that fool their respective discriminators. The cycle-consistency losses ($\mathcal{L}_{cycle,F}$ and $\mathcal{L}_{cycle,Q}$) enforce that a sample translated from one domain and then back to the original domain remains structurally consistent, thereby preserving anatomical fidelity. The identity losses ($\mathcal{L}_{iden,F}$ and $\mathcal{L}_{iden,Q}$) constrain each generator to behave as an identity mapping when the input already belongs to its target domain, preventing over-modification of high-quality inputs.
• Figure 5: Noise2Score / AR-DAE training flow: noisy image $y$ is perturbed by $u\!\sim\!\mathcal{N}(0,I)$ with scale $\sigma_a$ and fed to $R_\Theta$. The residual output is trained to provide the score needed for Tweedie’s denoising at test time.