Adaptive multi-resolution hash-encoding framework for INR-based dental CBCT reconstruction with truncated FOV

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Radiology

Background:

• Implicit Neural Representation (INR) with hash encoding shows promise for CT reconstruction.
• Applying INR to 3D dental cone-beam CT (CBCT) with truncated fields of view (FOV) is challenging due to projection discrepancies.
• Truncation artifacts degrade image quality in CBCT reconstruction.

Purpose of the Study:

• To develop a computationally efficient INR-based reconstruction framework for 3D dental CBCT with truncated FOV.
• To mitigate artifacts caused by truncated FOV in CBCT imaging.
• To leverage multi-resolution hash encoding for improved reconstruction.

Main Methods:

• Training the INR network over an expanded reconstruction domain encompassing the entire head.
• Employing an adaptive training strategy with a multi-resolution grid for computational efficiency.
• Introducing an adaptive hash encoder to manage varying resolutions and maintain network input dimensionality.

Main Results:

• The proposed method effectively mitigates truncation artifacts by using an extended FOV.
• The adaptive strategy reduces computational time by 60% compared to naive domain extension.
• Peak signal-to-noise ratio (PSNR) within the truncated FOV is preserved.

Conclusions:

• A novel INR-based reconstruction framework is presented for 3D dental CBCT with truncated FOV.
• The framework successfully reduces truncation artifacts and training costs.
• This approach offers an efficient solution for challenging CBCT reconstruction scenarios.