Computed tomography image reconstruction using stacked U-Net.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning, particularly convolutional neural networks, has shown promise in various image quality improvement tasks.
• Medical imaging faces challenges in acquiring sufficient training data due to patient privacy concerns.
• Existing X-ray computed tomography (CT) reconstruction methods can be time-consuming and may yield suboptimal image quality.

Purpose of the Study:

• To apply a deep learning method, stacked U-Net, for efficient and high-quality X-ray CT image reconstruction.
• To address the limitations of limited medical training data by leveraging external image databases.
• To reduce the number of projections required for accurate CT image reconstruction.

Main Methods:

• Utilized stacked U-Net, a deep learning architecture, for the image reconstruction task.
• Employed diverse images from the ImageNet database for model training to overcome medical data scarcity.
• Implemented the method for reconstructing 512x512 X-ray CT images using 64 projections, 512 detectors, and 360-degree rotation.

Main Results:

• Achieved a peak signal-to-noise ratio (PSNR) of 27.93 dB and a structural similarity index measure (SSIM) of 0.886 for the reconstructed images.
• Demonstrated a significantly reduced reconstruction time of 0.11 seconds on a GPU.
• The method produced superior image quality and faster processing compared to existing techniques.

Conclusions:

• The stacked U-Net deep learning model offers a promising approach for rapid and high-quality X-ray CT image reconstruction.
• Leveraging large external datasets like ImageNet can mitigate challenges associated with limited medical imaging data.
• This advancement has the potential to improve the efficiency and diagnostic capabilities of CT imaging.