Abstract
Cone-Beam Computed Tomography (CBCT) is a vital tool in dental and orthopedic imaging, offering rapid scans at lower radiation doses than conventional CT. However, reducing the number of X-ray projections to further minimize dose leads to significant "streak artifacts" that can obscure anatomical details. This paper presents a novel deep learning framework based on Generative Adversarial Networks (GANs) to perform artifact reduction as a post-processing task. By integrating a multi-scale Feature Pyramid Network (FPN) discriminator and a specialized "focus map," the proposed method reconstructs high-quality images from just one-third of the standard projections, maintaining structural integrity and diagnostic utility.
Introduction
Sparse-view CBCT is highly desirable for reducing patient radiation exposure and accelerating clinical workflows. The primary obstacle is the emergence of radiating streak artifacts during reconstruction from limited data. While traditional iterative methods attempt to solve this mathematically, they are often slow and computationally intensive. This work treats artifact reduction as an image-to-image translation problem, leveraging adversarial learning to map artifact-laden sparse-view images to clean, dense-view equivalents.
Methodology
The proposed framework consists of three primary components designed to address the multi-scale nature of streak artifacts and the challenge of localized noise.
Network Architecture
The generator utilizes an encoder-decoder architecture with skip connections (similar to a U-Net) to preserve fine anatomical details. To effectively penalize artifacts of varying lengths and thicknesses, the authors employ a multi-scale discriminator strategy. Specifically, a Feature Pyramid Network (FPN) discriminator evaluates the generated images at multiple semantic levels, ensuring both global consistency and local realism.
Differentially Modulated Focus Map
A core innovation of this research is the "focus map" ($\Lambda$). In standard GAN training, regions that are already relatively clean may produce ambiguous scores that do not drive effective learning. The focus map acts as a spatial weight mask derived from feature differences, forcing the generator to concentrate its optimization efforts on the most stubborn artifact-heavy regions.
Loss Functions
The model is trained using a combination of Least Squares GAN (LSGAN) loss for improved training stability and Perceptual Loss (via a pre-trained VGG16). This combination ensures that the output is not only mathematically accurate but also perceptually similar to real medical images, avoiding the blurring effects common with standard pixel-wise loss functions.
Experimental Results
The method was validated using a clinical dataset of knee CBCT scans from 27 subjects. Sparse-view inputs were reconstructed from 67 projections, while 200 projections provided the ground truth.
Quantitative Performance
The "Ours-Focus+FPN" model outperformed several baseline approaches, achieving a Structural Similarity Index (SSIM) of 0.884 and a Peak Signal-to-Noise Ratio (PSNR) of 36.38 dB. These metrics indicate a high degree of fidelity to the ground truth dense-view scans.
Qualitative and Computational Efficiency
Visual analysis showed that the proposed method effectively suppressed streak artifacts while preserving sharp bone boundaries. Computationally, the 2D version of the model processed an entire CBCT volume in approximately 16 seconds on a mid-range GPU, demonstrating its potential for near-real-time clinical integration.
Conclusion
This work introduces an effective adversarial framework for improving the quality of sparse-view CBCT. By utilizing multi-scale discrimination and a targeted focus map, the model successfully reduces streak artifacts without sacrificing anatomical detail. Future research will focus on extending the architecture to native 3D processing and conducting formal diagnostic evaluations with clinical experts.
Summary updated