Image reconstruction using UNET-transformer network for fast and low-dose PET scans.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Science

Background:

• Low-dose and fast Positron Emission Tomography (PET) imaging is vital for patient safety and comfort.
• Reconstructing high-quality images from low-count PET data is challenging due to noise and limited data.
• Effective denoising and image enhancement are critical for accurate PET scan interpretation.

Purpose of the Study:

• To develop an accurate deep learning-based method for low-count PET image reconstruction.
• To improve the quality of reconstructed PET images by reducing noise and preserving fine details.
• To combine the strengths of UNET and Transformer networks for enhanced PET image reconstruction.

Main Methods:

• The TrUNET-MAPEM model integrates a UNET-transformer regularizer with the maximum a posteriori expectation maximization (MAPEM) algorithm.
• A loss function combining structural similarity index (SSIM) and mean squared error (MSE) was used for accuracy evaluation.
• Performance was assessed using simulated (Brainweb phantom) and real patient data (Siemens Biograph mMR), compared against OSEM, MAPOSEM, and 3D-UNET.

Main Results:

• TrUNET-MAPEM achieved superior performance on both simulated and real patient data compared to state-of-the-art methods.
• For patient data, the model yielded an average PSNR of 33.72 dB, SSIM of 0.955, and rRMSE of 0.39.
• The model successfully reconstructed smooth images while preserving essential features like edges, outperforming other methods across all evaluated metrics.

Conclusions:

• The TrUNET-MAPEM model represents a significant advancement in low-count PET image reconstruction.
• The approach offers potential clinical applications for early disease detection and diagnosis.
• Results indicate superior noise reduction and edge preservation compared to existing algorithms.