HiResPET: high resolution PET image generation using deep convolution encoder decoder network

Aim: PET images are affected by a number of resolution degrading factors such as detector sampling width, inter-crystal scattering and penetration, and also positron range and photon non-collinearity. In the current study, we aimed to restore PET high-resolution images from a single low-resolution input PET image using a deep convolution encoder-decoder network.

Methods: 140 patients with brain PET images were included in the current study, and the data were divided into 100, 20 and 20 training, test, and external validation sets, respectively. Original ground-truth high-resolution images were used to simulate resolution degraded PET images via appropriate resolution degradation modeling in both image and projection space. Our HiResPET architecture consisted of convolutional encoder and decoder networks for end-to-end learning, invoking a U-NET like architecture. The network mapped low-resolution PET images to pixel-wise continuously-valued original high-resolution PET images via an encoder-decoder architecture. The encoder network performs efficient image compression while searching for robust and spatially invariant image features, and the decoder network reconstructs the desired output from the encoder output. Quality of the synthesized images were quantitatively assessed using root mean square error (RSME), peak signal-to-noise ratio (PSNR), and structural Similarity index metrics (SSIM). Image quantification was assessed using SUV bias map and joint histogram of pixel-wise SUV correlation between generated images and reference (ground truth high-resolution PET image and HiResPET). Result: With respect to reference high resolution PET images, RSME, PSNR and SSIM values were 0.0017± 0.0005 [0.0005-0.0045], 55.47±2.73 [46.91-64.62] and 0.999±0.0001 [0.997-0.999] for test set, and 0.0016±0.0003 [0.0004-0.0043], 56.47±1.45 [45.82-65.78] and 0.999±0.0001[0.998-0.999] for external validation set. Relative error (%) of SUV was 0.11±0.01 and 0.10±0.02 for the test set and external validation set, respectively. Pixel wise SUV correlation (Pearson correlation, R²) between generated images and reference was 0.99± 0.01 and 0.99± 0.02 for the test and external validation sets, respectively. Conclusion: In the current study, we aimed to generate high-resolution PET images using deep convolution encoder-decoder network. Our results from experiments on test and external validation sets (with 20 subjects for each) shows that our proposed method had high performance in recovering the images details. The proposed deep learning method shows significant promise for generating PET high -resolution images with important applications.

• Download figure
• Open in new tab
• Download powerpoint