Does Prior Knowledge in the Form of Multiple Low-dose PET Images (at Different Dose Levels) Improve Standard-dose PET Prediction?

Abstract

To acquire a high-quality PET image, a standard dose of radioactive tracer is injected into the patient which may pose a high risk of radiation exposure damage. On the other hand, reducing the injected dose increases the statistical noise within the PET images. To improve the image quality of low-dose PET (L-PET) images, deep learning methods have been introduced to denoise the L-PET images, wherein the relationship between the L-PET images and the standard-dose PET (S-PET) images is learned by the model to predict the SPET images from their low-dose counterparts. The existing deep learning-based approaches solely focus on a single level of L-PET imaging to predict the S-PET images. In this work, we investigate the benefits of exploiting multiple PET images at lower dose levels (in addition to the target low-dose level) as prior knowledge to predict the S-PET images. To this end, a high-resolution residual deep learning network was employed to develop two S-PET prediction models. First, the network was trained using a single input channel for 8% L-PET images. In the second model, multiple L-PET images (6% and 4%, in addition to 8% L-PET) were considered as inputs to the network. The performance of the two models was evaluated using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), root mean square error (RMSE) of standard uptake value (SUV) within the entire head region. Moreover, mean SUV bias (SUV <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mean</inf> ) was calculated for the malignant lesions. The quantitative analysis of 20 patients in the external validation dataset demonstrated the superior performance of the multi-input model. The RMSE within the entire head region reduced from 0.12±0.04 in 8% L-PET, to 0.09±0.03 and 0.06±0.02 for the single- and multi-input models, respectively. Moreover, the SUV <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mean</inf> bias reduced from -4.18±1.14% in the single-input model to -1.44±0.56% in the multi-input model. This study demonstrated the benefits of using multiple L-PET images to estimate the S-PET images.