Low-count PET image reconstruction based on truncated inverse radon layer and U-shaped network

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Positron Emission Tomography (PET) is crucial for diagnostics like tumor detection.
• Traditional PET image reconstruction is computationally intensive and struggles with low signal-to-noise ratios, especially at low radiation doses.
• Existing iterative reconstruction methods are time-consuming and may yield suboptimal image quality.

Purpose of the Study:

• To develop a deep learning-based method for direct PET image reconstruction from low-count sinograms.
• To improve image quality and reduce reconstruction time in low-dose PET imaging.
• To offer an alternative to conventional and current deep learning reconstruction techniques.

Main Methods:

• A novel deep learning network architecture was designed.
• The network incorporates a truncated inverse radon layer for domain transformation.
• A U-shaped network was utilized for image enhancement and reconstruction.

Main Results:

• The method was successfully validated on both simulated and real PET data.
• Structural similarity improved from 0.9357 to 0.9613 compared to ordered subset expectation maximization with Gaussian filtering.
• Peak signal-to-noise ratio showed a significant improvement of 5 dB.

Conclusions:

• The proposed deep learning method enables direct reconstruction of PET images from low-count sinograms.
• It achieves superior image quality and reduced processing time compared to iterative algorithms and state-of-the-art convolutional neural networks.
• This approach holds promise for improving efficiency and diagnostic accuracy in low-dose PET imaging.