Emis2Trans: Attenuation Correction for Brain PET With Many Types of PET Ligands Using Convolutional Neural Networks

Abstract

Attenuation correction is imperative in positron emission tomography (PET) measurements. A traditional transmission scan with an external positron line source has been replaced by X-ray CT or MRI. However, the scanner’s mechanics increases cost and these additional scans sometimes accompany image displacement which makes final images degraded in quality. Therefore, in this study, to overcome these problems, we propose an attenuation correction framework that generates transmission images from uncorrected emission images for brain PET imaging using deep convolutional neural networks (CNN). The advantage of this study is that an estimated transmission image is applicable in PET measurements with various types of ligands. Using encoder– decoder-based CNNs, the transmission scans for attenuation correction are generated from the uncorrected emission scans. We used the brain PET datasets from 1030 patients scanned with the following PET ligands: 18F-FDG, 18F-BCPP-EF, 11C-Racropride, 11C-PIB, 11C-DPA-713, and 11C-PBB3. We randomly selected 20% of the datasets as the testing dataset, and used the remaining 80% to train the network. As a result, the CNN generated was less noisy and had more uniform transmission scans compared to the original transmission scans (using a 68Ge-68Ga rotation line source). The CNN showed the peak signal-to-noise ratio and structural similarity indices as being 31.5 ± 1.8 dB and 0.80 ± 0.02 in generated and original transmission images of many types of PET ligands, respectively. These results indicate that the proposed CNN-based framework allows accurate attenuation correction in brain PET system using different types of PET ligands.