Ring Artifacts Removal Based on Implicit Neural Representation of Sinogram Data

Ligen Shi,Xu Jiang,YunZe Liu,Chang Liu,Ping Yang,Shifeng Guo,Xing Zhao

2024-09-25

Abstract:Inconsistent responses of X-ray detector elements lead to stripe artifacts in the sinogram data, which manifest as ring artifacts in the reconstructed CT images, severely degrading image quality. This paper proposes a method for correcting stripe artifacts in the sinogram data. The proposed method leverages implicit neural representation (INR) to correct defective pixel response values using implicit continuous functions and simultaneously learns stripe features in the angular direction of the sinogram data. These two components are combined within an optimization constraint framework, achieving unsupervised iterative correction of stripe artifacts in the projection domain. Experimental results demonstrate that the proposed method significantly outperforms current state-of-the-art techniques in removing ring artifacts while maintaining the clarity of CT images.

Image and Video Processing

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the problem of ring artifacts in CT imaging due to inconsistent responses of detector elements. Specifically, the inconsistent responses of X - ray detector elements can lead to streak artifacts in sinogram data, and these streaks appear as ring - like structures in the reconstructed CT images, seriously affecting the image quality. Therefore, the paper proposes a method based on the Implicit Neural Representation (INR) of sinogram data to correct the streak artifacts in the sinogram, thereby improving the quality of CT images.

### Problem Background
1. **Causes of Ring Artifacts**:
- The detector elements have inconsistent or damaged responses to X - rays.
- This inconsistency leads to data deviations at different viewing angles, thus forming streaks in the sinogram.
- During the CT image reconstruction process, these streaks will be back - projected into lines equidistant from the rotation center and accumulate at multiple viewing angles to form ring - like structures, that is, ring artifacts.

2. **Limitations of Existing Methods**:
- **Detector Calibration Methods**: Such as dark - field correction, flat - field correction, and bad - pixel correction. Equipment aging may lead to deviations in correction coefficients, making it difficult to completely remove ring artifacts.
- **CT Image Post - Processing Methods**: Involving Cartesian and polar coordinate transformations, which may lose image resolution.
- **Dual - Domain Iterative Methods**: Establish an optimization function by combining the characteristics of the projection domain and the CT image domain, but the streak artifacts appear as high - frequency information in the sinogram, which not only forms ring artifacts in the reconstructed image but also propagates throughout the image space, introducing additional noise and artifacts.
- **Data - Driven Methods**: Depend on the size and quality of the labeled dataset, and the lack of high - quality datasets in practical applications limits their effectiveness.

### The Method Proposed in the Paper
1. **Method Overview**:
- Use the Implicit Neural Representation (INR) to parameterize sinogram data and decompose it into an ideal sinogram (IS) and streak artifacts (SA).
- Through an unsupervised iterative optimization framework, simultaneously learn streak characteristics and correct the response values of defective pixels.
- Introduce a one - way regularization framework to perform local smoothing and sparse constraints on IS and SA respectively to achieve effective separation.

2. **Technical Details**:
- **Pixel Classification**: Classify the pixels in the projection data into defective pixels (caused by non - responsive detector elements) and non - defective pixels (caused by inconsistent responses of X - ray detectors).
- **Forward Model**: Establish a decomposition model of IS and SA and introduce effective regularization constraints.
- **Loss Function Design**: Combine the mean - squared absolute error and two regularization constraint terms.
- **Residual Compensation Strategy**: Compensate for the residual information after the decomposition of IS and SA to enhance image details.
- **Network Model Selection**: Select a network model that combines a multi - resolution grid module and a multi - layer perceptron (MLP) to parameterize IS, and select a learnable parameter matrix to represent SA.

### Experimental Results
The paper verifies the effectiveness of the proposed method through simulation data and real - data experiments. The experimental results show that this method can remove ring artifacts while maintaining the clarity of the image, which is significantly better than the existing advanced methods.

### Main Contributions
1. Propose an unsupervised streak - artifact - removal method that does not rely on labeled datasets.
2. Use the continuous function fitted by INR - parameterized IS and use the data of non - defective pixels to predict the projection values of defective pixels.
3. Introduce a one - way regularization framework to effectively separate IS and SA and improve the effect of streak - artifact removal.

Through these contributions, this paper provides a new and effective solution for ring - artifact removal in CT imaging.