Streak artefact removal in x-ray dark-field computed tomography using a convolutional neural network

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence

Background:

• X-ray computed tomography (CT) has limitations with weakly attenuating organs like the lung.
• X-ray dark-field (DF) imaging offers complementary structural information via small-angle scattering.
• Human-scale DF CT development faces challenges with streak artifacts impacting image interpretation.

Purpose of the Study:

• To demonstrate the feasibility of using CNNs for artifact reduction in DF CT.
• To establish multi-rotation scans as a method for generating training data.
• To improve image quality for potential clinical adoption of DF CT.

Main Methods:

• A supervised deep learning approach using a 3D dual-frame UNet was employed.
• Training data was generated using an experimental DF CT prototype.
• Input data used clinically relevant, dose-compatible scans; ground truth used extended, low-artifact scans.

Main Results:

• The trained CNN significantly reduced streak artifacts in DF CT images.
• Image quality metrics were quantitatively improved.
• Fine details were preserved, though output images were smoother than ground truth.

Conclusions:

• CNNs show significant potential for reducing streak artifacts in X-ray DF CT.
• Enhanced image quality in dose-compatible DF CT is crucial for clinical radiology adoption.