PET Imaging Super-Resolution Using Attention-Enhanced Global Residual Dense Network

Positron emission tomography (PET) is frequently utilized in various clinical applications, such as cancer diagnosis, heart disease screening, and neurological illness diagnosis. PET image super-resolution (SR) seeks to obtain clinically useful PET pictures of the highest quality at the lowest possible cost and patient risk. In this study, we offer the attention-enhanced global residual dense network (AGRDN) model of convolutional neural networks (CNNs) for small-animal PET image-to-image super-resolution. To stop low-level feature deterioration, AGRDN uses dense and skip connections in a recursive structure as the feature extractor. We carried out both accurate simulation experiments utilizing the phantom datasets, small animal PET datasets and ADNID to validate the AGRDN structures. The trials show that the peak signal-to-noise ratio (PSNR), structural similarity index measurement (SSIM), root mean square error (RMSE), and contrast-to-noise ratio (CNR) can all be improved using the AGRDN. Especially on the phantom datasets, other models are inferior to low-resolution images in PSNE, SSIM and RMSE, and our proposed model exceeds 0.49dB in PSNR.