Cone beam computed tomography reconstruction from truncated projections using prior information and transfer learning

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medical Imaging

Background:

• Cone beam computed tomography (CBCT) is essential for image guidance in clinical and research settings.
• Truncation artifacts severely degrade CBCT reconstruction quality when the field of view (FOV) is insufficient.

Purpose of the Study:

• To develop a Dual-Domain Deep learning-based method for CBCT Reconstruction from Truncated projections (D3CRT).
• To leverage non-truncated prior information to guide the reconstruction process and mitigate truncation artifacts.

Main Methods:

• D3CRT employs a dual-domain approach, combining projection and image domain processing.
• A Sinogram Generation Network (SG-Net) predicts missing projection data, followed by FDK reconstruction.
• An Image Enhancement Network (IE-Net) refines images, incorporating compressed sensing for data fidelity.

Main Results:

• D3CRT significantly improved reconstruction quality for truncated projections, outperforming low-resolution prior images.
• Achieved high Dice Similarity Coefficients (DSC) of 97.1% (whole-body) and 96.0% (lung regions).
• Demonstrated superior performance with lower RMSE (2.95 × 10⁻³ mm⁻¹) and higher SSIM (98.1%) compared to LRICR.

Conclusions:

• The D3CRT method effectively enhances CBCT reconstruction quality from truncated projections.
• Leveraging non-truncated prior information and projection-domain transfer learning is key to D3CRT's success.
• This approach offers a promising solution for improving image quality in undersampled CBCT acquisitions.