A novel convolutional neural network for predicting full dose from low dose PET scans

The use of radiolabeled tracers in PET imaging raises concerns owing to potential risks from radiation exposure. Therefore, to reduce this potential risk in diagnostic PET imaging, efforts have been made to decrease the amount of radiotracer administered to the patient. However, decreasing the injected activity reduces the signal-to-noise Ratio (SNR) and deteriorates image quality, thus adversely impacting clinical diagnosis. Previously proposed techniques are complicated and slow, yet they yield satisfactory results at significantly low dose. In this work, we propose a deep learning algorithm to reconstruct full-dose (FD) from low-dose (LD) PET images using a fully convolutional encoder-decoder deep neural network model. The goal is to train a model to learn to reconstruct from images with only 5% of the counts to produce images corresponding to 100% of the dose. Brain PET/CT images of 140 patients acquired on the Siemens Biograph mCT with a standard injected activity of F-18-FDG (205 +/- 10 MBq). Images were acquired for about 20 min. The sinograms of each scan were used to produce a low-dose sinogram by randomly selecting only 1/20(th) of the counts. To avoid over fitting, data augmentation was used. A modified 3D U-Net, was developed to predict standard-dose sinogram (PSS) from their corresponding LD sinogram. Detailed quantitative and qualitative comparison demonstrated the proposed method can generate artefact-free diagnostic quality images that preserve internal structures without noise amplification. The structural similarity index (SSIM) and peak signal to noise ratio (PSNR) were used as quantitative metrics for assessment. For instance, the PSNR and SSIM in selected slices were 37.30 +/- 0.71 and 0.97 +/- 0.02, respectively. The proposed algorithm operates in the projection space and is capable of producing diagnostic quality images with only 5% of the standard injected activity.