OSCNet: Orientation-Shared Convolutional Network for CT Metal Artifact Learning.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Metal artifacts in X-ray computed tomography (CT) hinder accurate disease diagnosis and image-guided interventions.
• Existing deep learning methods for metal artifact reduction (MAR) often fail to fully utilize the physical characteristics of these artifacts.

Purpose of the Study:

• To develop a novel deep learning framework for effective metal artifact reduction (MAR) in CT imaging.
• To incorporate prior knowledge of artifact characteristics into the reconstruction process for improved performance.

Main Methods:

• Proposed an orientation-shared convolution representation mechanism to model the rotationally symmetrical patterns of metal artifacts.
• Utilized Fourier-series-expansion-based filter parametrization for artifact modeling and separation from body tissues.
• Developed the orientation-shared convolutional network (OSCNet) using deep unfolding techniques.
• Introduced a dynamic convolution sub-network for improved artifact learning flexibility, leading to OSCNet+.

Main Results:

• OSCNet and OSCNet+ demonstrated significant effectiveness in reducing metal artifacts on both synthetic and clinical CT datasets.
• The proposed methods successfully separated metal artifacts from body tissues, enhancing image quality.
• OSCNet+ showed improved generalization performance due to its flexible artifact learning capabilities.

Conclusions:

• The proposed orientation-shared convolution mechanism and dynamic convolution representation are effective for metal artifact reduction in CT.
• OSCNet and OSCNet+ offer a promising solution for improving the quality and reliability of CT imaging in the presence of metallic implants.