Ring Artifacts Correction Based on Global-Local Features Interaction Guidance in the Projection Domain

Yunze Liu, Congyi Su, Xing Zhao · 2025

Read Ring Artifacts Correction Based on Global-Local Features Interaction Guidance in the Projection Domain by Yunze Liu, Congyi Su, Xing Zhao (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-11375.txt.

Introduction

Yunze Liu, Congyi Su, and Xing Zhao propose a method to remove ring artifacts from CT scans. Their 2025 paper starts in the projection data, before images are rebuilt. Rings arise when detector units respond unevenly to X-rays, leaving stripes that become concentric rings in the final image.

Older methods treat those stripes as fixed errors. The authors show that detector unevenness scales with the projection values themselves, so it shifts during a scan. Their method joins global features, which should vary smoothly, with local features, where stripes stand out sharply. Guiding each by the other helps pull stripes out without harming true structure. Cleaner scans matter for diagnosis, since rings can hide or mimic real features in the body.

Catalog record

CREATORYunze Liu, Congyi Su, Xing Zhao
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-11375

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Ring Artifacts Correction Based on Global-Local Features Interaction Guidance in the Projection Domain Yunze Liu, Congyi Su, Xing Zhao Abstract Ring artifacts are common artifacts in CT imaging, typically caused by inconsistent responses of detector units to X-rays, resulting in stripe artifacts in the projection data. Under circular scanning mode, such artifacts manifest as concentric rings radiating from the center of rotation, severely degrading image quality. In the Radon transform domain, even if the object’s density function is piecewise discontinuous in certain regions, the projection images remain nearly continuous in the angular direction, making the ideal projections exhibit a smooth global low-frequency characteristic. In practical scanning, the local disturbances of the same detector unit at different scanning angles lead to a prominent high-frequency locality of stripe artifacts. Existing studies generally model ring artifacts disturbances as fixed additive errors, which overlooks the dynamic variation of detector responses during practical scanning. However, the degree of detector response inconsistency is a function of the projection values, as revealed in our experiments, thereby requiring consideration of the interaction between global and local features in the process of stripe artifacts extraction and correction. Therefore, we propose a CT ring artifacts correction method based on global and local features in the projection domain. We employ the VSS block and Dense block to respectively correct the low-frequency sub-band, which capture the global correlations of the projection, and the high-frequency sub-band, which contain local stripe artifacts after wavelet decomposition. Specifically, the accuracy of artifacts correction is enhanced by the interaction guidance between global and local features. Extensive experiments demonstrate that our method achieves superior performance in both quantitative metrics and visual quality, verifying its robustness and practical applicability. I Introduction Ring artifacts in CT images mainly arise from the inconsistent responses of detector units to X-ray. These inconsistencies caused by factors such as variations in scintillation screen thickness, non-uniform detector sensitivity, or amplifier circuit fluctuations [ 1 ] . These inconsistencies lead to significant deviations in recorded values even under identical X-ray exposure. When projection data are arranged in angular order, such errors appear as stripe artifacts along the detector direction, as illustrated in the projection domain of Fig. 1 . After reconstruction, these stripe artifacts appear as multiple concentric rings centered at the image center, with increasing intensity toward the center, as shown in the image domain of Fig. 1 . These artifacts obscure structural details and compromise the accuracy of lesion diagnosis, defect detection, and material analysis in CT imaging. Therefore, effective correction of ring artifacts has become a critical topic for improving CT image quality. Fig. 1: Characterization of ring artifacts in CT image and projection domains. Currently, ring artifacts correction methods in CT imaging can be broadly categorized into image-domain, projection-domain, and dual-domain methods. Image-domain methods typically transform ring artifacts into stripe artifacts via polar coordinate conversion for easier correction, followed by inverse transformation to the Cartesian domain. Representative methods include segmentation-based [ 2 , 3 , 4 ] , transform-based [ 5 , 6 ] , decomposition-based [ 7 , 8 , 9 ] , and total variation (TV)-based techniques [ 10 , 11 , 12 , 13 , 14 ] . However, interpolation errors induced by coordinate conversion often degrade image resolution and introduce secondary artifacts. In recent years, deep learning-based image-domain correction methods have emerged. For instance, Hein et al. employed a U-Net architecture with spectral-VGG loss [ 15 ] , while Zhao et al. developed Smooth GAN using adversarial learning with smoothness constraints [ 16 ] . Nevertheless, these methods are limited by insufficient generalization, structural distortion, and detail loss. Projection-domain methods often utilize frequency decomposition for artifacts suppression. A typical example is the Wavelet-Fourier Filtering (WFF) method [ 17 ] , where Discrete Wavelet Transform (DWT) is used for high-frequency and low-frequency separation, followed by frequency-domain filtering. Guo et al. further improved this method by introducing weighted averaging and damping filters [ 18 ] . Deep learning has also been introduced for projection-domain correction. Nauwynck et al. proposed a U-Net-based network with a weighted loss leveraging projection sparsity [ 19 ] . Fu et al. combined residual learning and transfer learning, using wavelet subbands to separate artifacts while preserving details [ 20 ] . Dual-domain methods combine both image and projection priors within an optimization framework [ 21 , 22 , 23 ] . Moreover, Fang et al. proposed a five-stage U-Net to perform domain-specific correction, integrating image, projection, and polar domain processing with parallel training for improved artifact suppression [ 24 ] . The Radon transform, proposed by Johann Radon in 1917, provides the mathematical foundation for projection acquisition and image reconstruction in CT imaging [ 25 , 26 ] . It describes the process of acquiring projection data by integrating a two-dimensional image along a set of specified lines. Moreover, Wu et al. [ 27 ] demonstrated that the Radon transform remains continuous for almost all scanning angles, even when the object’s density function exhibits boundary or jump discontinuities. This theoretical result indicates that the ideal projection data should exhibit well smoothness and global consistency. Fig. 2: (a) Illustration of the inverse discrete wavelet reconstruction using the vertical high-frequency sub-band and the low-frequency sub-band from a real projection image. (b) Profile curve along the red, blue, and green lines shown in (a). In fan-beam CT scanning, detector response inconsistency arises from both intrinsic hardware variation and differences in ray paths through the object. This leads to angle-dependent variations in the response of a single detector unit, resulting in localized stripe artifacts in the projection data. To model this effect, several studies [ 22 , 12 , 13 ] approximate the corrupted projection as the sum of the ideal projection and a fixed additive error term, which can be formulated as: P^{\prime}_{i}=P_{i}+G_{i} (1) where P_{i} denotes the ideal projection, P^{\prime}_{i} represents the corrupted projection, and G_{i} is the additive error term associated with the i -th detector unit. In this case, the projection image with stripe artifacts is modeled as the sum of a constant additive term G_{i} and the ideal projection P_{i} . However, this formulation neglects the angle-dependent variability of detector response, limiting the effectiveness of additive-model-based correction methods. To investigate this variability of a single detector unit, we applied DWT to a measured projection image P_{\text{measured}} , obtaining a low-frequency sub-band image P^{LL}_{\text{measured}} and three high-frequency sub-band images in horizontal, vertical, and diagonal directions: P^{LH}_{\text{measured}} , P^{HL}_{\text{measured}} , and P^{HH}_{\text{measured}} , respectively. Owing to the geometric characteristics of stripe artifacts, DWT facilitates their localization and separation. As shown in Fig. 2 (a), the vertical high-frequency sub-band effectively captures the localized features of stripe artifacts, while the low-frequency sub-band retains the global structure of the projection image. We further reconstructed the low-frequency and high-frequency sub-band images separately to obtain a global image P_{\text{measured}}^{G} and a local image P_{\text{measured}}^{L} . Then, we extracted the values of the same detector unit across different projection angles from P_{\text{measured}} , P^{G}_{\text{measured}} , and P^{L}_{\text{measured}} . These responses are annotated with red, green, and blue lines, respectively in Fig. 2 (a), and their corresponding value curves are plotted in Fig. 2 (b). The results indicate that the inconsistency in detector response

Read the complete text