Standard-Dose PET Prediction from Low-Dose Versions: Investigation of the Post-Reconstruction Denoising Filters as Prior Knowledge in a Deep Learning Solution

Low-dose PET (L-PET) imaging, in which the injected radiopharmaceutical dose has been reduced to decrease the risk of radiation exposure damage, would result in PET images with poor image quality due to the increased statistical noise. To improve the quality of L-PET images, various studies have been carried out using deep learning methods in which standard-dose PET (S-PET) images have been predicted from their low-dose counterparts. The existing deep-learning based approaches mainly focus on applying single input of L-PET images to the deep learning networks. To the best of our knowledge, no studies so far have been conducted to employ multiple inputs (different versions of the same L-PET image) to the deep learning models as additional information. In this work, in order to predict S-PET images, we utilized filtered L-PET images using various post-reconstruction filters (including the Gaussian, bilateral, self-guided, anisotropic diffusion, non-local means, and median filters) as the input of the deep-learning network, in addition to the original L-PET image to assess the benefits of applying additional information obtained from filtered images. To this end, a high-resolution Resnet network was employed/developed to implement seven independent S-PET prediction models. The first model was trained using only the original 3% L-PET images. For the six remaining models, the filtered L-PET images using the aforementioned filters in addition to the original L-PET images were considered as input to the networks. The quality of the predicted images from the different models was evaluated using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), root mean square error (RMSE) of standard uptake value (SUV) in the entire head region and malignant lesions. Moreover, the bias of max SUV (SUVmax bias) in the predicated S-PET images with respect to the reference S-PET images for the malignant lesions was calculated. The quantitative assessment of these metrics across 40 patients indicated the superior performance of the network trained with L-PET images filtered with self-guided filter in addition to the original L-PET images to the other networks. The RMSE within the entire head region, and the lesion SUVmax bias reduced by 1.36% and 67.27% comparing to the single-input model (regarded as conventional fashion model training), respectively. Furthermore, training the model using self-guided filtered images achieved higher values of SSIM and PSNR compared to the models. The investigation of the different approaches to predict S-PET images from the corresponding L-PET images using seven independent deep learning models indicated the superiority of the network trained with multiple inputs including the original L-PET image and the L-PET images filtered with the self-guided filter compared to the other filters and single input model.