Unsupervised Domain Adaptation for Low-dose CT Reconstruction via Bayesian Uncertainty Alignment

TL;DR

LDCT reconstruction under real-world domain shifts suffers from degraded target-domain performance and lack of uncertainty awareness. The authors introduce a probabilistic cross-domain framework with a Bayesian encoder–decoder and two complementary alignment mechanisms: Bayesian uncertainty alignment (BUA) in the latent space to reduce epistemic gap via covariance-based latent statistics and a self-reconstruction constraint, and sharpness-aware distribution alignment (SDA) in the image space to match second-order image information using a maximum local variation descriptor within an adversarial setting. The approach optimizes a joint objective $L = \beta_1 L_{SL} + \beta_2 L_{BUA} + \beta_3 L_{SDA}$ and leverages MC sampling with mean-field variational inference to quantify uncertainty and drive domain-invariant representations. Experiments on simulated (AAPM-16, AAPM-A, AAPM-B) and clinical (ISIDCM-20) LDCT datasets show consistent improvements in both objective metrics (PSNR, SSIM, GMSD, DSS) and qualitative texture/edge fidelity, along with favorable uncertainty maps that support reliability in unseen acquisitions.

Abstract

Low-dose computed tomography (LDCT) image reconstruction techniques can reduce patient radiation exposure while maintaining acceptable imaging quality. Deep learning is widely used in this problem, but the performance of testing data (a.k.a. target domain) is often degraded in clinical scenarios due to the variations that were not encountered in training data (a.k.a. source domain). Unsupervised domain adaptation (UDA) of LDCT reconstruction has been proposed to solve this problem through distribution alignment. However, existing UDA methods fail to explore the usage of uncertainty quantification, which is crucial for reliable intelligent medical systems in clinical scenarios with unexpected variations. Moreover, existing direct alignment for different patients would lead to content mismatch issues. To address these issues, we propose to leverage a probabilistic reconstruction framework to conduct a joint discrepancy minimization between source and target domains in both the latent and image spaces. In the latent space, we devise a Bayesian uncertainty alignment to reduce the epistemic gap between the two domains. This approach reduces the uncertainty level of target domain data, making it more likely to render well-reconstructed results on target domains. In the image space, we propose a sharpness-aware distribution alignment to achieve a match of second-order information, which can ensure that the reconstructed images from the target domain have similar sharpness to normal-dose CT images from the source domain. Experimental results on two simulated datasets and one clinical low-dose imaging dataset show that our proposed method outperforms other methods in quantitative and visualized performance.

Figures (11)

• Figure 1: Left: Uncertainty quantification on source and target domains using BNN-based reconstruction model with DA strategy (Model w/ DA trained by paired data from the source domain and LDCT images from the target domain) and BNN-based reconstruction model without DA strategy (Model w/o DA trained by source domain data only). A higher level of uncertainty is observed on the target domain for the model without DA strategy. Besides, the level of uncertainty on the target domain is close to the source domain by adopting our proposed DA method. Right: Examples of an LDCT image from the source domain, an LDCT image from the target domain, and corresponding reconstructed results on target domain data. The display window is [-160,240] HU.
• Figure 2: The overall framework of the proposed method. The CT reconstruction model is decoupled into a BNN-based encoder for the feature extraction and a BNN-based decoder for the content reconstruction from extracted features. (a) Bayesian uncertainty alignment module (in the latent space): Uncertainty discrepancy term $\mathcal{L}_{UD}$ aims to explicitly quantify the uncertainty and reduce this discrepancy between source and target domains, as described in Section \ref{['udml']}; Self-reconstruction term $\mathcal{L}_{SR}$ as a complementary constraint encourages extracting invertible latent features for the target domain in the process of uncertainty discrepancy minimization, as described in section \ref{['srl']}. (b) Sharpness-aware distribution alignment module (in the image space): This module encourages the model to reconstruct high-quality CT images for target domains with a similar level of sharpness (as second-order information) as the NDCT images from the source domain, as described in section \ref{['Sharpness-aware Distribution']}.
• Figure 3: (a): An example NDCT from source domain; (b) An example LDCT from target domain; (c) NDCT version of (b); (d), (e) and (f) is the visualized results of MLV map for (a), (b) and (c). (g), (i) (h) and (j) are a series of histogram comparisons of different images in the first two rows. Comparison between (g) and (i) implies that there is still an obvious distribution discrepancy even if the LDCT images (b) on the target domain have been reconstructed well on (c), due to a mixture of content and style information. Instead, (h) implies that the sharpness-based histogram distribution is well separable. (j) show a high distribution consistency if the LDCT images (b) on the target domain have been reconstructed well.
• Figure 4: Qualitative results of an example CT image from AAPM-A dataset. The red ROI is zoomed in for visual comparison, and the red arrow points to one lesion. The absolute differences between the reconstructed results and corresponding NDCT images NDCT are shown in the second and last row. The display window is [-160,240] HU.
• Figure 5: Qualitative results of an example CT image from AAPM-B dataset. The red ROI is zoomed in for visual comparison, and the red arrow points to one lesion. The absolute differences between the reconstructed results and corresponding NDCT images NDCT are shown in the second and last row. The display window is [-160,240] HU.