[Paper Review] Ring Artifact Correction in Photon-Counting Spectral CT Using a Convolutional Neural Network With Spectral Loss

Dennis Hein, Konstantinos Liappis+3 more|arXiv (Cornell University)|Feb 2, 2023

Advanced X-ray and CT Imaging4 citations

TL;DR

This paper proposes a deep learning-based post-processing method for ring artifact correction in photon-counting spectral CT using a UNet with a novel spectral loss combining L1 loss on material basis images and perceptual loss on 70 keV virtual monoenergetic images. The method effectively reduces ring artifacts while preserving fine anatomical details, producing high-quality 40, 70, and 100 keV virtual monoenergetic images without requiring real-time detector calibration.

ABSTRACT

Photon-counting spectral computed tomography is now clinically available. These new detectors come with the promise of higher contrast-to-noise ratio and spatial resolution and improved low-dose imaging. However, one important design consideration is to build detector elements that are sufficiently homogeneous. In practice, there will always be a degree of inhomogeneity in the detector elements, which will materialize as variations in the energy bin thresholds. Without proper detector calibration, this will lead to streak artifacts in the sinograms and corresponding ring artifacts in the reconstructed images, which limit their clinical usefulness. Since detector calibration is a time-consuming process, having access to a fast ring correction technique may greatly improve workflow. In this paper, we propose a deep learning-based post-processing technique for ring artifact correction in photon-counting spectral CT. We train a UNet with a custom loss to correct for ring artifacts in the material basis images. Our proposed loss is made ``task-aware'' by explicitly incorporating the fact that we are working with spectral CT by combining a L1-loss operating on the material basis images with a perceptual loss, using VGG16 as feature extractor, operating on 70 keV virtual monoenergetic images. Our results indicate that using this novel loss greatly improves performance. We demonstrate that our proposed method can successfully produce ring corrected 40, 70, and 100 keV virtual monoenergetic images.

Motivation & Objective

• To address ring artifacts in photon-counting spectral CT caused by detector inhomogeneity and energy threshold variations.
• To develop a fast, post-processing alternative to time-consuming detector calibration.
• To improve image quality in spectral CT by correcting ring artifacts while preserving diagnostic details.
• To explore the effectiveness of a task-aware loss function combining L1 and perceptual losses for spectral CT image restoration.
• To evaluate the method on clinical-relevant virtual monoenergetic images across multiple energy levels.

Proposed method

• A U-Net architecture is trained to correct ring artifacts directly in the image domain on material basis images.
• A custom spectral loss function combines a pixel-wise L1 loss on material basis images with a perceptual loss on 70 keV virtual monoenergetic images.
• The perceptual loss uses features from a pre-trained VGG16 network to preserve texture and structural details.
• The network is trained on 512×512 patches extracted from clinical CT scans of the chest and abdomen.
• The method operates in the image domain without coordinate transformation to polar space, simplifying implementation.
• Training is performed end-to-end using a combination of L1 and perceptual losses to balance fidelity and perceptual quality.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively correct ring artifacts in photon-counting spectral CT without relying on sinogram-domain processing?
• RQ2 Does a spectral loss combining L1 and perceptual losses improve artifact correction while preserving fine anatomical details?
• RQ3 How well does the method generalize to different energy levels (40, 70, 100 keV) in virtual monoenergetic images?
• RQ4 Can the model handle ring artifacts in challenging anatomical regions like the head, where current training data is limited?
• RQ5 Is the proposed spectral loss more effective than standard L1 or L2 losses for spectral CT ring correction?

Key findings

• The proposed method successfully produces ring-corrected 40, 70, and 100 keV virtual monoenergetic images with minimal residual artifacts.
• The use of a spectral loss combining L1 and perceptual losses significantly improves image quality compared to standard loss functions.
• The network preserves fine details and textures, avoiding over-smoothing common in L1-only approaches.
• The method performs well on torso scans (chest and abdomen), but struggles with ring artifacts in head scans due to lack of training data in this anatomical region.
• Initial experiments with adversarial loss (WGAN-GP) did not significantly improve performance and were abandoned due to high computational cost.
• Larger patch sizes (512×512) were found to be necessary for generalization in spectral CT, as smaller patches (e.g., 128×128) led to poor performance.