Using LSO background radiation for CT-less attenuation correction of PET data in long axial FOV PET scanners

Objectives: The present generation of PET scanners use lutetium-based scintillators due to their short decay time, high density, and high luminosity. The radioisotope ¹⁷⁶Lu in these scintillators provides intrinsic background radiation that can be used to measure attenuation without an additional dose from CT. This is of particular interest in view of the recent introduction of long axial FOV scanners, which are suitable for extremely low dose PET for pediatric imaging, repeated scans, PET screening, and more. Moreover, due to the larger number of detectors, the rate of such background events in an extended axial FOV scanner is far greater than that of conventional PET scanners. In this study, a novel CT-less PET image reconstruction framework using LSO background transmission (LSO-TX) data is discussed and evaluated on clinical PET/CT scanners.

Methods: The Biograph Vision and Biograph Vision Quadra (Siemens Healthineers) PET/CT scanners were used in this work. The Biograph Vision and Vision Quadra have, respectively, 8 and 32 detector rings providing an axial coverage of 26 and 106 cm in a single FOV. Detectors of both scanners comprise LSO crystals with an estimated ¹⁷⁶Lu activity of 240 Bq/cm³. With a wide-open energy window, a detector singles rate of about 11 kcps is recorded due to the LSO intrinsic activity. In a ¹⁷⁶Lu decay, a β decay with a Q value of 597 keV followed by a cascade of γ photons at 307 and 202 keV occurs. To capture the 202 and 307 keV γ photons of LSO TX, the energy and coincidence timing window were wide open. While the β triggers the constant fraction discriminator, the γ photon records a coincidence. From the list of recorded events, the LSO-TX at 202 and 307 keV were extracted and histogrammed to create transmission sinograms, and μ-maps were reconstructed using a maximum likelihood for transmission tomography (MLTR) algorithm. The μ-maps from LSO-TX suffer from noise and often have a low signal-to-noise ratio (SNR). Therefore, a convolutional neural network (CNN) based image denoising method was employed. U-Net architecture with residual units was used to denoise LSO-TX images. 2D axial images of LSO-TX-derived μ-maps were used as input data. The image volumes were normalized to zero mean and unity variance. The network was trained on 9 various phantom datasets acquired using Biograph Vision and Biograph Vision Quadra. The resulting μ-maps were fed into a TOF-MLAA algorithm in which activity and attenuation images were sequentially updated using, respectively, the TOF-MLEM and MLTR algorithms. The extremely good time resolution of the two scanners allows for optimal performance of TOF-MLAA.

Results: Several experiments were performed to develop and validate the method in both Biograph Vision and Biograph Vision Quadra. In this summary, MLAA images of a NEMA NU-2 image quality phantom with a total of 15 MBq ⁶⁸Ge acquired on Biograph Vision are shown. Also, an example of LSO TX-derived μ-map is shown for a human subject, as acquired on the Biograph Vision and Biograph Vision Quadra. Other phantom and patient images from the Vision Quadra will be shown in the final presentation. LSO-TX-derived μ-maps with and without CNN-based denoising are shown in figure 1 for the image quality phantom. Compared to CT-derived μ-map, mean absolute error of initial LSO TX μ-maps was reduced from 0.028 cm^-1 to 0.016 cm^-1 when CNN-based denoising is applied.

Conclusions: PET images reconstructed from the TOF MLAA algorithm initialized with denoised LSO TX μ-maps, and TOF MLEM algorithm using CT-based μ-maps suggest that the TOF MLAA reconstruction framework can be a reliable and suitable solution to CT-less PET imaging, in particular for long axial FOV PET scanners.

• Download figure
• Open in new tab
• Download powerpoint