Utilizing Pix2Pix conditional generative adversarial networks to recover missing data in preclinical PET scanner...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Gaps in Positron Emission Tomography (PET) detector blocks cause data loss, degrading image quality and quantitative accuracy.
• This data loss impacts reconstructed PET images, particularly in preclinical applications.

Purpose of the Study:

• To develop and evaluate a novel approach for filling missing sinogram data in preclinical PET scanners.
• To improve the image quality and quantitative accuracy of reconstructed PET images affected by detector block gaps.

Main Methods:

• Utilized a Pix2Pix conditional generative adversarial network (cGAN) combined with an inpainting technique.
• Trained the network on 7500 raw sinograms from preclinical PET scans of mice and an IQ phantom.
• Artificially generated gap-free sinograms for training due to the absence of real gap-free data.

Main Results:

• The Pix2Pix cGAN achieved a low Root Mean Squared Error (RMSE) of 9.34 × 10⁻⁴ ± 5.7 × 10⁻⁵.
• Demonstrated a high Structural Similarity Index (SSIM) of 99.984 × 10⁻² ± 1.8 × 10⁻⁵, indicating excellent image fidelity.
• Quantitative metrics like PSNR, CNR, and SNR were used for performance assessment.

Conclusions:

• The proposed Pix2Pix cGAN effectively retrieves missing sinogram data by learning from adjacent pixel information.
• This method significantly enhances quantitative accuracy and improves the quality of reconstructed PET images.
• The approach offers a viable solution for mitigating data loss issues in preclinical PET imaging.