Trans²-CBCT: A Dual-Transformer Framework for Sparse-View CBCT Reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Sparse-view Cone-Beam Computed Tomography (CBCT) reduces radiation dose and scan time.
• However, sparse-view CBCT suffers from severe streak artifacts and spatial coverage gaps.
• Existing reconstruction methods struggle to address these limitations effectively.

Purpose of the Study:

• To develop a unified framework for improving sparse-view CBCT reconstruction.
• To enhance image quality by mitigating artifacts and improving spatial coverage.
• To leverage advanced deep learning architectures for superior CBCT image reconstruction.

Main Methods:

• Introduced Trans-CBCT using a hybrid CNN-Transformer (TransUNet) encoder for local and long-range spatial context modeling.
• Developed Trans²-CBCT by incorporating a neighbor-aware Point Transformer to enforce volumetric coherence.
• Utilized multi-scale feature maps and an attenuation-prediction head for CBCT reconstruction.

Main Results:

• Trans-CBCT outperformed baseline methods by 1.17 dB in PSNR and 0.0163 in SSIM on the LUNA16 dataset with six projection views.
• Trans²-CBCT achieved further improvements, with an additional 0.63 dB in PSNR and 0.0117 in SSIM over Trans-CBCT.
• Both Trans-CBCT and Trans²-CBCT consistently outperformed prior methods on LUNA16 and the ToothFairy dataset across multiple metrics.

Conclusions:

• Hybrid CNN-Transformer architectures effectively model local and global features for sparse-view CBCT.
• Geometry-aware point-based reasoning enhances volumetric coherence and reconstruction accuracy.
• The proposed Trans-CBCT and Trans²-CBCT frameworks offer significant advancements for low-dose, sparse-view CBCT imaging.