PSFNET: ultrafast generation of PSF-modelled-like PET images using deep convolutional neural network

Aim: PET imaging suffers from a limited spatial resolution and the resulting partial volume effect (PVE). Implementation of PSF modeling within iterative reconstruction (accounting for the point spread function of the PET system) attempts to compensate for the degradation of spatial resolution and to reduce PVE. We aimed to generate PSF-modelled-like higher-resolution PET images using deep convolutional neural network, as an ultrafast alternative to conventional PSF-modeled image reconstructions.

Methods: Ninety-one patient’s data with brain PET were used in this study. Data were randomly partitioned into 70, 10 and 11 patients for the training, test, and external validation sets, respectively. PSFNET consists of encoder and decoder networks where the encoder converts non-PSF images to feature vectors and the decoder reconstructs the PSF-like image with end to end learning. Quality of the synthesized images were quantitatively assessed by mean absolute error (MAE), 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 (PSFNET) and reference (PSF images).

Results: With respect to reference original PSF images MAE, RMSE, PSNR and SSIM values were 0.0036±0.0015 [0.0008-0.010], 0.012±0.0043 [0.004-0.032], 38.23±2.71 [29.78-46.63], 0.992±0.008 [0.89-0.99] and 0.0033±0.0011 [0.0007-0.012], 0.011±0.0041 [0.004-0.029], 39.21±3.41 [32.48-47.13], 0.994±0.003 [0.92-0.99] for the test set and external validation set. Pixel wise SUV correlation (Pearson correlation, R²) between PSFNET images and reference PSF image was 0.99± 0.01 and 0.98± 0.05 for the test set and external validation set. Relative error (%) of SUV was -5.12±1.01 [-6.02-1.4] and -5.02±2.13 [-6.52-2.1] for the test set and external validation set respectively.

Conclusions: This study has identified the possibility of generating PSF-modelled-like PET images using deep convolutional neural network modeling. The proposed deep convolutional neural network method shows promise for generation of high-resolution PET images in the context of improved quantitative analysis.

• Download figure
• Open in new tab
• Download powerpoint