Abstract
In this work, we tackle the problem of robust computed tomography (CT) reconstruction under a cross-domain scenario, where training data (source domain) and testing data (target domain) are collected from different anatomical regions. Due to mismatches in scan regions and protocols, noise distribution shifts often lead to a catastrophic deterioration of reconstruction performance on the target domain. To achieve robust performance, we introduce a Bayesian-endowed probabilistic framework. Under this framework, we propose to alleviate noise distribution shifts via implicit noise modeling in both the latent and image spaces. Specifically, a novel Bayesian Noise Uncertainty Alignment (BNUA) method is proposed for the latent space, while an adversarial Residual Distribution Alignment (RDA) reduces discrepancies in the image space. Extensive experiments on head and abdomen scans demonstrate that our method achieves superior robust cross-domain CT reconstruction compared to existing approaches in both quantitative and qualitative terms.
Introduction
Low-dose CT (LDCT) imaging is a critical clinical tool for reducing patient radiation exposure. However, the resulting images are significantly noisier than standard-dose scans. While deep learning models have shown great promise in denoising these images, they are often "brittle" when faced with domain shift. A model trained on a specific dataset (the source domain) frequently fails when deployed in a clinic using different scanners or protocols (the target domain).
Existing Unsupervised Domain Adaptation (UDA) solutions often struggle with content mismatch—where anatomical differences between patients confuse the model—and a lack of reliability, as deterministic models cannot quantify their own uncertainty. This paper addresses these challenges by proposing a probabilistic reconstruction framework that leverages uncertainty as a bridge between domains.
Methodology
The proposed framework utilizes Bayesian Neural Networks (BNNs) to treat model weights as distributions rather than fixed values. This allows the system to quantify epistemic uncertainty, or the "gap" in the model's knowledge when encountering unfamiliar data.
As shown in Figure 4, the framework incorporates two primary alignment mechanisms:
- Bayesian Noise Uncertainty Alignment (BNUA): In the latent feature space, the model calculates the covariance of multiple Monte Carlo samples. By minimizing the difference between the uncertainty levels of the source and target domains, the model learns to extract domain-invariant features.
- Residual Distribution Alignment (RDA): To handle image-space noise without distorting anatomy, the authors employ a Sharpness-aware Distribution Alignment (SDA) scheme. This uses Maximum Local Variation (MLV) descriptors to extract "sharpness" maps that capture noise and texture style while ignoring specific anatomical structures. An adversarial discriminator then ensures the target reconstructions match the sharpness style of high-quality source scans.
Experiments and Results
The study utilized four distinct datasets to simulate domain shift: AAPM-16 (source), AAPM-A (heavy noise target), AAPM-B (mild noise target), and ISICDM-20 (real-world clinical target from a different manufacturer).
Quantitative Performance
The proposed method consistently outperformed state-of-the-art UDA methods like CycleGAN and CCDnet. On the heavy-noise AAPM-A task, the framework achieved a PSNR of 32.68 dB, a significant improvement over standard UDA approaches (30.87 dB) and non-adapted backbones (28.56 dB).
Qualitative and Clinical Impact
Visual results (Figure 4b) demonstrate that the alignment of noise uncertainty leads to much clearer reconstructions with fewer artifacts. In blind studies, professional radiologists rated the method highest for structure preservation and overall clinical usability. Notably, in real-world clinical datasets (ISICDM-20), the model maintained tiny lesions that other methods either smoothed over or distorted, providing a reliable "trust meter" through its generated uncertainty maps.
Conclusion
By integrating Bayesian uncertainty quantification with dual-space noise alignment, this framework provides a robust solution for deploying LDCT reconstruction models across diverse clinical environments. Future work will explore more sophisticated probability modeling, such as Normalizing Flows, to further enhance reconstruction accuracy in highly challenging cross-domain cases.
Summary updated