PET Image Reconstruction using Convolutional Neural Network and Generative Adversarial Network in Sinogram Domain

Kupcikevicius, Ignas

4. term

Biomedical Engineering and Informatics, Master

2019

2019-06-06

10 pages

Background: Artifacts and noise in PET imaging are caused by multiple factors, including lowered radiotracer dose and information loss in the form of missing pixels or missing projections. Recently, deep learning based algorithms have achieved promising results in the medical imaging field, including PET denoising or CT sinogram inpainting, especially using the Convolutional Neural Network (CNN) and Generative Adversarial Networks (GAN) architectures. This article aims to compare CNN and GAN approaches for PET sinogram missing data reconstruction task. Methods: The end-to-end framework, from PET image to the sinogram domain and back to PET image domain, was proposed. The Radon transform was applied to covert PET images into sinograms. The first model was the CNN encoder-decoder based network with four skip connections. The effective strategy was applied to efficiently train more corrupted PET sinograms by loading previously trained weights. The second approach was the GAN network, with the generator designed similarly as the CNN encoder-decoder, and the discriminator containing four convolutional layers to classify generated sinograms as artificially generated or ground truth. The proposed framework ended by applying filter back projection algorithm to transform sinograms back to PET image domain. Results: The results revealed that GAN outperformed CNN by a small margin. The average PSNR and SSIM scores within all five corruption levels were 41.44, 0.977, and 42.34, 0.983 when predicting missing pixels. Differences of two metrics between CNN and GAN were higher when predicting missing projections; 40.13, 0.866 versus 46.84, 0.989. Additionally, GAN performed noticeably better when 90% of sinogram data were removed, resulting in a sharper and more detailed reconstructed image, qualitatively comparing to CNN. Conclusion: Different network architectures, chosen, and objective functions, might be the reasons why GAN performed better than CNN. Even though the study had some limitations, the promising results were achieved, which motivates to experiment further.

Background: Artifacts and noise in PET imaging are caused by multiple factors, including lowered radiotracer dose and information loss in the form of missing pixels or missing projections. Recently, deep learning based algorithms have achieved promising results in the medical imaging field, including PET denoising or CT sinogram inpainting, especially using the Convolutional Neural Network (CNN) and Generative Adversarial Networks (GAN) architectures. This article aims to compare CNN and GAN approaches for PET sinogram missing data reconstruction task. Methods: The end-to-end framework, from PET image to the sinogram domain and back to PET image domain, was proposed. The Radon transform was applied to covert PET images into sinograms. The first model was the CNN encoder-decoder based network with four skip connections. The effective strategy was applied to efficiently train more corrupted PET sinograms by loading previously trained weights. The second approach was the GAN network, with the generator designed similarly as the CNN encoder-decoder, and the discriminator containing four convolutional layers to classify generated sinograms as artificially generated or ground truth. The proposed framework ended by applying filter back projection algorithm to transform sinograms back to PET image domain. Results: The results revealed that GAN outperformed CNN by a small margin. The average PSNR and SSIM scores within all five corruption levels were 41.44, 0.977, and 42.34, 0.983 when predicting missing pixels. Differences of two metrics between CNN and GAN were higher when predicting missing projections; 40.13, 0.866 versus 46.84, 0.989. Additionally, GAN performed noticeably better when 90% of sinogram data were removed, resulting in a sharper and more detailed reconstructed image, qualitatively comparing to CNN. Conclusion: Different network architectures, chosen, and objective functions, might be the reasons why GAN performed better than CNN. Even though the study had some limitations, the promising results were achieved, which motivates to experiment further.

PET reconstruction ; Sinogram domain ; CNN ; FBP ; GAN ; Radon transform

Download
View record in AAU Student Projects