Self-supervised denoising of projection data for low-dose cone-beam CT

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Image Denoising

Background:

• Convolutional Neural Networks (CNNs) show promise for image denoising.
• Supervised CNNs require high-quality references, which are often unavailable in interventional radiology like cone-beam computed tomography (CBCT).
• Existing methods struggle with the lack of clean targets for CBCT denoising.

Purpose of the Study:

• To propose a novel self-supervised learning method for reducing noise in CBCT projections.
• To enable high-quality CBCT image reconstruction from low-dose projections without requiring reference data.

Main Methods:

• A self-supervised learning network partially blinds input projections to train denoising models.
• Noise-to-noise learning is incorporated by mapping adjacent projections to original projections.
• Denoised projections are reconstructed into CBCT images using standard algorithms like FDK.

Main Results:

• The self-supervised denoising approach achieved significantly higher Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) values compared to uncorrected low-dose CBCT.
• Phantom studies showed PSNR of 27.08 and SSIM of 0.839 for the proposed method, versus 15.68 and 0.103 for uncorrected data.
• Retrospective studies confirmed the effective production of high-quality CBCT images from low-dose projections, validated both qualitatively and quantitatively.

Conclusions:

• The proposed self-supervised learning strategy effectively removes noise from CBCT projection data.
• Anatomical information is successfully restored in denoised CBCT images.
• This method offers a viable solution for low-dose CBCT imaging in interventional radiology.