Technical Summary: Progressive J-Invariant Self-supervised Learning for Low-Dose CT Denoising

Problem Statement

Low-dose computed tomography (LDCT) is critical for reducing radiation exposure in clinical settings such as cancer screening and pulmonary nodule detection. However, the reduced radiation dose inherently introduces significant noise and artifacts, compromising image clarity and diagnostic confidence. While supervised deep learning methods (e.g., U-Net, RED-CNN) have achieved strong denoising performance, they rely heavily on paired normal-dose CT (NDCT) and LDCT data, which are often scarce or impossible to acquire due to ethical and radiation constraints.

Existing self-supervised alternatives, particularly "blind-spot" methods like Noise2Void and Noise2Self, attempt to eliminate the need for clean targets by enforcing conditional independence (J-invariance), where a pixel's prediction is independent of its own noisy observation. However, these conventional approaches suffer from two primary limitations:

1. Training Inefficiency: They typically impose conditional independence in a single step with restricted receptive fields, leading to suboptimal learning of structural details.
2. Overfitting and Instability: Without clean references, these models are prone to overfitting the noise or failing to learn robust contextual features, especially given the complex, non-Gaussian, and spatially correlated noise characteristics of LDCT (which differs from the additive noise models often assumed in natural image denoising).

Methodology

The authors propose Progressive J-Invariant Learning (PJSL), a novel self-supervised framework designed to enhance LDCT denoising by maximizing the utility of the J-invariant property through a multi-stage training process.

Core Mechanism

The method reframes the denoising task as a temporally progressive process rather than a single-step mapping.

• Step-wise Blind-Spot Denoising: Instead of a single forward pass, the network operates over $K$ time steps. At each step $t$, the denoised output $\hat{x}_t$ serves as the input for the next step $\hat{x}_{t+1}$.
• Progressive Masking: To enforce J-invariance, a random binary mask $M$ is applied to the input at each time step. This mask is applied to either the raw LDCT input or the intermediate denoised output. This strategy treats progressive denoising stages as "gradually-amplified data augmentation," allowing the network to incrementally refine structural details and aggregate contextual information over multiple steps.
• Controlled Noise Injection: To prevent overfitting and improve the learning of pixel correlations, the authors inject controlled zero-mean additive white Gaussian noise (AWGN) into both the input and the target during training. Theoretical analysis suggests that minimizing the loss between noisy inputs and noisy targets (where noise is independent and zero-mean) indirectly minimizes the supervised loss between the network output and the clean ground truth.

Training and Inference

• Training: A single model is trained across different time steps with varying random masks. The loss function minimizes the difference between the masked, noise-perturbed input and the noise-perturbed target.
• Inference: During testing, $k$ differently masked versions of the same LDCT image are fed into the trained model. The final denoised image is obtained by averaging the corresponding outputs, leveraging the ensemble effect to stabilize the result.

Key Contributions

1. Progressive J-Invariant Strategy: The paper introduces a step-wise masking and denoising mechanism that overcomes the training inefficiencies of conventional blind-spot methods. By progressively enforcing conditional independence, the model learns to recover fine-grained anatomical structures more effectively.
2. Noise-Based Regularization: The authors explicitly inject controlled Gaussian noise into both inputs and targets. This serves as a regularization mechanism that mitigates overfitting and enhances the robustness of the self-supervised learning process without violating the J-invariant assumptions.
3. Architecture-Agnostic Performance: The proposed strategy is demonstrated to be compatible with various backbone architectures (U-Net, Swin-Unet, RED-CNN), providing consistent improvements across different network designs.
4. Theoretical Justification: The paper provides a theoretical derivation showing that the proposed noise injection and progressive loss formulation align with the goal of minimizing the supervised loss, even in the absence of clean NDCT targets.

Experimental Results

Extensive experiments were conducted on the Mayo LDCT Grand Challenge dataset, utilizing 27 patients for training and 9 for evaluation (ensuring subject-level separation).

• Quantitative Performance: The proposed method (PJSL) consistently outperformed existing self-supervised approaches (Noise2Void, Neighbor2Neighbor, BM3D) across all metrics (PSNR, SSIM, RMSE).
  • On the abdominal dataset, PJSL with a U-Net backbone achieved a PSNR of 31.510 dB, surpassing the supervised RED-CNN (31.748 dB) and significantly outperforming other self-supervised methods.
  • On the more challenging chest dataset, PJSL achieved a PSNR of 18.131 dB, outperforming the supervised RED-CNN (19.792 dB) and self-supervised baselines.
• Comparison with Supervised Methods: PJSL achieved performance comparable to, and in some cases superior to, representative supervised methods (e.g., CycleGAN, RED-CNN) without requiring paired NDCT data.
• Ablation Studies:
  • Time Steps ($k$): Performance improved as $k$ increased from 1 to 5, reaching an optimum at $k=5$. Further increases led to degradation due to error accumulation.
  • Noise Injection: Combining progressive masking with noise injection yielded the best results, confirming that noise injection prevents overfitting caused by high similarity between input and target.
  • Mask Ratio ($\alpha$): An optimal mask ratio of 0.1 was identified; lower ratios provided weak supervision, while higher ratios removed too much contextual information.

Significance and Claims

The paper claims that Progressive J-Invariant Learning significantly narrows the performance gap between self-supervised and supervised denoising methods without increasing model complexity. By addressing the training inefficiencies of traditional blind-spot networks and the overfitting issues inherent in self-supervised learning, the method enables high-quality LDCT denoising in scenarios where paired clean data is unavailable.

The authors emphasize that their approach relaxes the restrictive assumptions of Noise2Noise (requiring paired noisy observations) and the data scarcity constraints of supervised learning. They position the method as a practical solution for clinical environments where acquiring NDCT-LDCT pairs is infeasible, offering a robust, architecture-agnostic framework that balances noise removal with the preservation of critical anatomical details. The work concludes by noting limitations regarding error accumulation in deep progressive steps and the current restriction to the LDCT dataset, suggesting future extensions to other imaging modalities like LDPET.