A dual-stream deep convolutional network for reducing metal streak artifacts in CT images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Metal artifacts significantly degrade computed tomography (CT) image quality, posing challenges for accurate diagnosis and treatment planning.
• Existing metal artifact reduction (MAR) algorithms, including state-of-the-art methods, often struggle with residual artifacts, particularly in demanding clinical scenarios such as proton therapy.

Purpose of the Study:

• To develop and evaluate a novel dual-stream deep convolutional neural network for effective metal artifact reduction in CT images.
• To improve image quality for proton therapy planning by accurately delineating tumor volumes in the presence of metallic implants.

Main Methods:

• A dual-stream deep convolutional neural network architecture incorporating residual learning was trained for streak removal.
• The network utilized a mask of metal streaks to focus artifact correction and employed a dual-stream approach for local structure correction and attention-based destreaking.
• Experiments compared mean squared error loss with perceptual loss functions to assess their impact on image feature and texture preservation.

Main Results:

• The proposed dual-stream deep learning network demonstrated significant effectiveness in reducing metal streak artifacts.
• Both visual and quantitative assessments confirmed improved image quality in cases with metal implants.
• The dual-stream processing, combining local correction and attention mechanisms, contributed to successful artifact removal.

Conclusions:

• Image-domain deep learning presents a highly effective strategy for metal artifact reduction in CT imaging.
• The study highlights the benefits of dual-stream network architectures and the influence of different loss functions on artifact correction and image fidelity.