End-to-end deep learning for interior tomography with low-dose x-ray CT

Area of Science:

• Medical Imaging
• Radiology
• Computer Vision

Background:

• X-ray computed tomography (CT) employs strategies like sparse-view, low-dose, and region-of-interest (ROI) CT to reduce radiation dose.
• Combining these techniques can lead to coupled artifacts, such as cupping artifacts from truncated projections and noise from low-dose settings.
• Existing image-domain deep learning (DL) methods struggle to resolve these combined artifacts effectively.

Purpose of the Study:

• To develop a novel method for reconstructing high-quality CT images from data with coupled artifacts.
• To address the limitations of current DL approaches in handling combined low-dose and ROI CT challenges.

Main Methods:

• Decoupled the coupled artifact problem into two sub-problems: noise reduction (low-dose CT) and projection extrapolation (ROI CT).
• Developed a novel end-to-end learning method utilizing dual-domain Convolutional Neural Networks (CNNs).
• Implemented a projection-domain CNN to address specific artifact types.

Main Results:

• The proposed dual-domain CNN method significantly outperforms conventional image-domain DL techniques.
• A projection-domain CNN demonstrated superior performance compared to commonly used image-domain CNNs.
• The method effectively reduces radiation dose and mitigates cupping artifacts and noise in reconstructed CT images.

Conclusions:

• The proposed dual-domain DL approach offers a superior solution for CT image reconstruction with coupled artifacts.
• Decoupling the problem and utilizing dual domains (image and projection) enhances artifact correction capabilities.
• This method holds promise for improving diagnostic accuracy and patient safety in CT imaging.