An Uncertainty-guided Manifold Smoothing Method for Non-Ideal Measurement Computed Tomography Reconstruction

Fengzhi Xu, Ziyuan Yang, Zexin Lu, Yingyu Chen, FENGLEI FAN, Hongming Shan, Yi Zhang

Keywords: CT Reconstruction, Image Synthesis, Unsupervised Learning

Abstract: Non-ideal measurement computed tomography (NICT) reduces the need for extensive data sampling, accelerating scanning and mitigating radiation exposure risks. However, the reconstructed images often suffer from artifacts and noise. While enormous deep learning (DL) methods have been developed to improve image quality, most rely on paired data, which is challenging to obtain due to physiological motion. Unsupervised reconstruction methods is a possible solution to address this issue, but they typically assume homogeneous noise distributions and ignore distinct noise characteristics arising from different sampling strategies, which may cause model collapse under certain conditions. We observe that NICT images form discrete sub-manifolds in feature space due to varying physical scanning processes, which contradicts the assumption of unsupervised methods and consequently limits their effectiveness. To address this, we propose an Uncertainty-Guided Manifold Smoothing (UMS) framework to bridge the gaps between sub-manifolds. In UMS, a classifier is first trained to identify the sub-manifold associated with each feature representation. The predicted uncertainty scores are then used to guide the generation of diverse samples across the entire manifold. By leveraging the classifier’s capability, UMS effectively fills the gaps between discrete sub-manifolds, and promotes a more continuous and dense feature space. Furthermore, due to the complexity of the global manifold, it's hard to directly model the manifold. Therefore, we propose to dynamically incorporate the global- and sub-manifold-specific features. Specifically, we design a global- and sub-manifold-driven architecture guided by the classifier, which enables dynamic adaptation to subdomain variations. This dynamic mechanism improves the network’s capacity to capture both shared and domain-specific features, thereby improving reconstruction performance. Extensive experiments on the public datasets are conducted to validate the effectiveness and generalizability of our method.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 8857