Efficient high cone-angle artifact reduction in circular cone-beam CT using deep learning with geometry-aware...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology

Background:

• High cone-angle artifacts (HCAAs) are prevalent in circular cone-beam computed tomography (CBCT), impacting diagnostic accuracy and treatment planning.
• Existing artifact reduction methods may not fully address the complexities of 3D HCAAs in CBCT data.

Purpose of the Study:

• To propose and evaluate a novel deep learning approach for reducing HCAAs in CBCT images.
• To compare the proposed method against a traditional Cartesian slice-based deep learning technique.
• To assess the impact of artifact reduction on the accuracy of CBCT image segmentation.

Main Methods:

• A novel deep learning approach was developed, transforming 3D HCAA reduction into efficient 2D problems by exploiting rotational scanning geometry.
• Convolutional neural networks (CNNs), specifically U-Net and mixed-scale dense CNN (MS-D Net), were trained on radially sampled CBCT slices.
• Performance was evaluated against a Cartesian slice-based CNN approach, and the effect on image segmentation quality was analyzed.

Main Results:

• The proposed geometry-aware deep learning approach demonstrated significant reduction of HCAAs in CBCT scans.
• The novel method outperformed the Cartesian slice-based deep learning approach in artifact reduction.
• Artifact reduction using the proposed method markedly improved the accuracy of subsequent CBCT image segmentation.

Conclusions:

• The developed deep learning approach effectively reduces HCAAs in CBCT images by leveraging geometric information.
• This method offers superior performance compared to Cartesian slice-based techniques for HCAA removal.
• Improved artifact reduction leads to enhanced accuracy in downstream CBCT image segmentation tasks.