Enhancing the Image Quality via Transferred Deep Residual Learning of Coarse PET Sinograms

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence in Healthcare

Background:

• Improving positron emission tomography (PET) image quality is crucial for clinical applications.
• Thin-pixelated crystals offer high spatial resolution but compromise sensitivity and increase manufacturing costs.
• Large pixelated crystals present challenges like blurred sinograms and parallax errors, impacting image quality.

Purpose of the Study:

• To develop an approach for enhancing PET image resolution and noise properties in scanners utilizing large pixelated crystals.
• To overcome limitations of traditional thin-crystal approaches by leveraging advanced imaging techniques.

Main Methods:

• A data-driven, single-image super-resolution (SISR) method based on a deep residual convolutional neural network (CNN) was developed for sinogram enhancement.
• Specialized techniques including periodic padding of sinogram data and a dedicated network architecture were employed for PET imaging efficiency.
• Transfer learning was incorporated to handle scenarios with limited labeled data or small training datasets.

Main Results:

• The proposed SISR method successfully improved PET image resolution and noise characteristics when using large pixelated crystals.
• Achieved comparable image resolution to thin crystals (bin size 2.5 mm) while offering superior noise performance.
• Validation was performed using analytical simulations, Monte Carlo simulations, and pre-clinical data.

Conclusions:

• The developed deep learning approach effectively enhances PET image quality with large pixelated crystals, offering a viable alternative to expensive thin-crystal systems.
• The method seamlessly integrates into existing PET imaging frameworks and requires no additional information during inference.
• This technique holds potential for the design of cost-effective, high-performance PET systems.