DeepPET: A deep encoder-decoder network for directly solving the PET image reconstruction inverse problem.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiological Sciences

Background:

• Clinical positron emission tomography (PET) image reconstruction faces challenges with algorithm optimization and computational demands.
• Existing methods often lack automated optimization and are computationally expensive, hindering widespread adoption of advanced techniques.

Purpose of the Study:

• To develop and implement a deep learning network, DeepPET, for efficient and high-quality PET image reconstruction.
• To address bottlenecks in automated optimization and computational expense for advanced PET reconstruction algorithms.

Main Methods:

• A novel end-to-end deep convolutional encoder-decoder network (DeepPET) was designed.
• The network takes PET sinogram data as input and directly outputs reconstructed PET images.
• Extensive training and validation were performed using simulated data from a digital phantom, with over 291,000 augmented reference images.

Main Results:

• DeepPET demonstrated superior image quality compared to OSEM and FBP, with lower relative root mean squared error and higher structural similarity index and peak signal-to-noise ratio.
• Image reconstruction speed was significantly enhanced, being 108x faster than OSEM and 3x faster than FBP.
• The DeepPET network was successfully validated on real clinical PET data.

Conclusions:

• DeepPET offers a powerful end-to-end solution for PET image reconstruction, overcoming limitations of conventional methods.
• The deep learning approach significantly reduces reconstruction time while maintaining or improving image quality.
• This technique holds promise for advancing clinical PET imaging through faster, more accurate reconstructions.