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

Background: The presence of a gap between adjacent detector blocks in Positron Emission Tomography (PET) scanners introduces a partial loss of projection data, which can degrade the image quality and quantitative accuracy of reconstructed PET images. This study suggests a novel approach for filling missing data from sinograms generated from preclinical PET scanners using a combination of an inpainting method and the Pix2Pix conditional generative adversarial network (cGAN). Materials and methods: Twenty mice and Image Quality (IQ) phantom were scanned by a small animal Xtrim PET scanner, resulting in 7500 raw sinograms used for network training and test datasets. The absence of gap-free sinograms as the target for neural network training was a challenge. This issue was solved by artificially generating gap-free sinograms from the original sinogram. To assess the performance of the proposed approach, the sinograms were reconstructed using the ordered subset expectation maximization (OSEM) algorithm. The overall performance of the proposed network and the quality of the resulting images were quantitatively compared using various metrics, including the root mean squared error (RMSE), structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), contrast-to-noise ratio (CNR), and signal-to-noise ratio (SNR). Results: The Pix2Pix cGAN approach achieved an RMSE of 9.34 x 10-4 +/- 5.7 x 10-5 and an SSIM of 99.984 x 10-2 +/- 1.8 x 10-5, considering the corresponding inpainted sinograms as the target. Conclusion: The proposed approach can retrieve missing sinogram data by learning a map derived from the whole sinogram compared to the adjacent pixels, which leads to better quantitative accuracy and improved reconstructed images.