Artifact suppression for breast specimen imaging in micro CBCT using deep learning

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence

Background:

• Cone-beam computed tomography (CBCT) is used for breast-specimen imaging to ensure clear resection margins.
• Typical micro CT imaging faces challenges with long acquisition times and artifacts from reduced projections (streak artifacts) and metallic markers (metal artifacts).

Purpose of the Study:

• To develop and evaluate a deep learning-based approach for suppressing both streak and metal artifacts in CBCT breast specimen images.
• To improve the image quality and diagnostic utility of CBCT for breast conservation surgery.

Main Methods:

• A modified U-Net neural network was used to synthesize sinograms after removing metal objects and upsampling.
• The synthesized sinograms were reconstructed using filtered backprojection (FBP).
• A subsequent ResU-Net model was applied to further reduce residual artifacts, combining denoised images with extracted metal objects.

Main Results:

• The proposed deep learning method significantly reduced streak and metal artifacts compared to conventional FBP and iterative reconstruction.
• The method achieved a 3.6x higher contrast-to-noise ratio, 1.3x higher peak signal-to-noise ratio, and 1.4x higher SSIM.
• Improved visualization of soft tissues around metallic markers was observed.

Conclusions:

• The developed deep learning approach effectively reduces streak and metal artifacts in CBCT reconstructed breast specimen images.
• This enhancement in image quality holds significant potential for improving clinical decision-making in breast conservation therapy.