Efficient Noise Context-Aware Representation for Low-Dose CT Denoising

The gist The ENCORE framework proposes an Efficient Noise Context-aware Representation for Low-Dose CT Denoising by explicitly leveraging CT noise characteristics and anatomical features to enhance restoration performance and computational efficiency.

How it works

1. The framework reformulates the noise synthesis procedure based on the Cornish-Fisher expansion to achieve more accurate CT noise simulation compared to the conventional Gaussian approximation.

2. To provide more direct and stable context guidance than NADD’s raw stochastic noise augmentations, an efficient autocovariance estimation step is introduced to explicitly extract local noise correlation and power.

3. The FlyingConv module is proposed to adaptively change convolution weights for each local image region by fusing 4 × 4 average-pooled feature maps F and autocovariance maps V.

4. To enable zero-shot conditional denoising, the noise augmentation formulation is extended by scaling the estimated noise intensity from PLD according to the target dose level dtarget, allowing for dynamic control over the output image texture.

Key Components and Innovations

(The paper enumerates several key contributions)

(We reformulate a self-supervised training pipeline based on the Cornish-Fisher expansion, providing more realistic noise simulation compared to the conventional Gaussian approximation.)

(We propose noise autocovariance preprocessing for model-friendly noise context estimation, which minimizes computational overhead while boosting throughput several-fold.)

(We propose the FlyingConv module, which adaptively adjusts the weight values based on the local anatomy and noise context information, enhancing both denoising quality and zero-shot robustness with minimal inference latency overhead.)

Experimental Setup and Results

The experiments utilized the 2016 NIH-AAPM-Mayo Clinic LowDose CT Grand Challenge dataset (Mayo2016) for training and testing. The framework was integrated into non-hierarchical DnCNN [20] and hierarchical UNet [21] backbones, comparing Vanilla, +NADD, +COV, and +ENCORE configurations.

Qualitatively, the ENCORE configuration reconstructs fine structures with high fidelity and demonstrates superior capability in handling complex noise patterns compared to models that solely rely on the noisy LDCT input.

Quantitatively, Table I indicates that +ENCORE yields the best results across all metrics, showing a substantial reduction in MACs compared to other configurations.

The analysis of ablation studies showed that replacing the Gaussian approximated term with the skewness-corrected term W derived from the Cornish-Fisher expansion improved SSIM specifically in high-attenuation slices where severe photon starvation causes the conventional Gaussian approximation to fail.

Furthermore, zero-shot denoising evaluation revealed a clear trade-off between PSNR and AUHOC when sweeping the dtarget value, enabling dynamic modulation of output texture during inference without requiring any additional training or post-processing.

Conclusion on Performance Trade-offs

The framework demonstrates that integrating noise context serves as a more effective paradigm for low-dose CT denoising than simply expanding the denoising model. While the reduction in MACs is substantial, the actual inference latency remains comparable to that of the +COV variant because the FlyingConv module manipulates weight values dynamically, which increases memory access overhead and strains GPU memory bandwidth. This suggests that hardware-level constraints related to memory bandwidth are currently limiting further latency gains. The success on the real-world Tabletop dataset validates the robustness of the method under physical imaging conditions, despite relying on a simplified noise model.

Future Directions

In future work, accounting for missing physical factors such as scatter and beam hardening will help bridge the gap between simulation and real clinical data to lead to an even more faithful CT denoiser. The framework's ability to flexibly tune the balance between texture preservation and denoising strength through zero-shot control is highly valuable in self-supervised learning environments such as N2N. The study confirms that incorporating noise context profiles yields quality gains with only a minimal latency increase, highlighting the critical necessity for a model tailored specifically to CT noise contexts. The superior performance of +ENCORE over all alternative variants in most cases is confirmed by statistical testing (p < 0.01).

Benchmarking Efficiency

Customized CUDA kernels provide a remarkable throughput boost, with the FBP module achieving a 4.8× acceleration compared to LEAP-based FBP, while the covariance estimation and FlyingConv-based denoiser modules deliver 2.9× and 3.2× speedups over their naive PyTorch baselines. The group size g = 2 in the FlyingConv module achieves the optimal balance between computational efficiency and denoising performance, as demonstrated in Table IV.

Ablation Study on Context Estimation

Using w = 5 yields the best performance for the window size, outperforming the baseline (w = 1) which strictly adheres to the original autocovariance definition. Similarly, p = 5 yields the best results for patch size, as it spans an area that is neither too localized nor overly broad for 3×3 convolution kernels. The study concludes that setting N = 1 is sufficient to capture the noise context because additional FBP executions and increases the computational overhead do not yield noticeable performance gain.

Final Summary of Contributions

The primary contributions include reformulating a self-supervised training pipeline based on the Cornish-Fisher expansion for more realistic noise simulation. The introduction of noise autocovariance preprocessing for model-friendly noise context estimation minimizes computational overhead while boosting throughput. Finally, the FlyingConv module dynamically adjusts convolution weights based on local anatomy and noise context information, enhancing both denoising quality and zero-shot robustness with minimal inference latency overhead. The entire pipeline is available at https://github.com/minwoo-yu/ENCORE.git.

--- Page 1 ---

ENCORE: Efficient Noise Context-Aware Representation for Low-Dose CT Denoising Minwoo Yu, N. Robert Bennett, Jongduk Baek, and Adam S. Wang Abstract—While deep learning-based denoising has become widely adopted in low-dose CT, conventional models use generic architectures designed for natural images, failing to account for non-stationary and spatially correlated CT noise characteristics. To address this, we propose an Efficient Noise COntext-aware REpresentation (ENCORE) framework that explicitly leverages CT noise characteristics and anatomical features. First, we reformulate the noise synthesis procedure based on a realistic noise distribution beyond the conventional Gaussian approximation, establishing a rigorous foundation for training pair generation. Next, we extract local noise power and correlation contexts to guide the denoising process. To fully leverage the potential of noise context, we propose a FlyingConv module, which adaptively changes convolution weights for each local image region. Notably, our approach demonstrates substantial gains in both denoising quality and computational efficiency. Furthermore, manipulating the intensity of the noise context maps at inference time enables zero-shot conditional denoising, allowing for dynamic control over the output image texture. The entire pipeline is available at https://github.com/minwoo-yu/ENCORE.git.

1. The ENCORE framework can perform zero-shot conditional denoising by manipulating the intensity of the noise context maps at inference time enables zero-shot conditional denoising, allowing for dynamic control over the output image texture.

2. The system can be trained on self-supervised learning pipelines where noise augmentation is stochastic by nature but stabilized by reformulating noise synthesis using the Cornish-Fisher expansion to achieve more realistic CT noise simulation compared to the conventional Gaussian approximation.

3. The denoising model can adaptively adjust its operation via the FlyingConv module, which adaptively changes convolution weights for each local image region, enabling a balance between quality and efficiency by utilizing an optimized structure where the group size g = 2 achieves the optimal balance between computational efficiency and denoising performance.

4. The system can be optimized across the entire imaging chain by developing custom CUDA kernels, which have been reengineered specifically to maximize GPU cache efficiency, resulting in a 4.8× acceleration compared to LEAP-based FBP for reconstruction and speedups for the denoiser modules.

5. The system can dynamically adjust texture preservation during inference by using the zero-shot control mechanism, where setting dtarget = 1 matches the residual noise in the output image to the normal-dose level.