Better Cone-Beam CT Artifact Correction via Spatial and Channel Reconstruction Convolution Based on Unsupervised...

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Cone-Beam Computed Tomography (CBCT) is crucial for image-guided radiotherapy (IGRT).
• CBCT images, particularly of soft tissues, suffer from artifacts and noise, impacting diagnostic accuracy.
• Existing methods struggle to effectively correct these artifacts while preserving anatomical integrity.

Purpose of the Study:

• To develop and evaluate a new unsupervised algorithm for CBCT image artifact correction.
• To improve the quality, clarity, and realism of CBCT images.
• To reduce artifacts and preserve anatomical structures in CBCT scans.

Main Methods:

• Proposed a novel unsupervised algorithm named Spatial Convolution Diffusion (ScDiff).
• ScDiff utilizes a conditional diffusion model, integrating Generative Adversarial Networks (GANs) and diffusion model characteristics.
• Employed a combination of unsupervised learning and stable training for artifact correction.

Main Results:

• ScDiff efficiently and stably corrected CBCT image artifacts.
• The algorithm produced clear, realistic CBCT images with preserved anatomical structures.
• ScDiff outperformed several GAN- and diffusion-based methods in corrected image quality and evaluation metrics.

Conclusions:

• ScDiff effectively enhances CBCT image quality by reducing artifacts.
• The proposed method preserves crucial anatomical details, aiding in accurate diagnosis.
• ScDiff represents a significant advancement in CBCT image artifact correction for IGRT applications.