A geometry-guided deep learning technique for CBCT reconstruction.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning (DL) for computed tomography (CT) reconstruction faces memory challenges in cone-beam CT (CBCT).
• Sparse data acquisition in CBCT exacerbates reconstruction difficulties.

Discussion:

• A novel geometry-guided deep learning (GDL) technique is proposed, integrating a GDL reconstruction module and a DL post-processing module.
• The GDL reconstruction module utilizes an array of small fully connected layers, adapting to projection geometry to overcome memory limitations.
• Performance is evaluated against traditional methods like Feldkamp, Davis, and Kress (FDK) and ray-tracing, with and without DL post-processing.

Key Insights:

• GDL significantly improves CBCT image quality, as evidenced by enhanced peak-signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and reduced root-mean-square error (RMSE).
• The GDL technique requires four orders of magnitude less memory compared to existing DL methods, making DL-based CBCT reconstruction practical.
• Reconstruction times are comparable across methods, highlighting GDL's efficiency.

Outlook:

• GDL represents a breakthrough, enabling rapid and accurate DL-based CBCT reconstruction from sparsely sampled data.
• This advancement holds potential for improving diagnostic accuracy and efficiency in various medical imaging applications.
• Further research could explore GDL's application in other low-data or memory-constrained imaging modalities.