Obtaining PET/CT images from non-attenuation corrected PET images in a single PET system using Wasserstein generative adversarial networks

Abstract

Positron emission tomography (PET) imaging plays an indispensable role in early disease detection and postoperative patient staging diagnosis. However, PET imaging requires not only additional computed tomography (CT) imaging to provide detailed anatomical information but also attenuation correction (AC) maps calculated from CT images for precise PET quantification, which inevitably demands that patients undergo additional doses of ionizing radiation. To reduce the radiation dose and simultaneously obtain high-quality PET/CT images, in this work, we present an alternative based on deep learning that can estimate synthetic attenuation corrected PET (sAC PET) and synthetic CT (sCT) images from non-attenuation corrected PET (NAC PET) scans for whole-body PET/CT imaging. Our model consists of two stages: the first stage removes noise and artefacts in the NAC PET images to generate sAC PET images, and the second stage synthesizes CT images from the sAC PET images obtained in the first stage. Both stages employ the same deep Wasserstein generative adversarial network and identical loss functions, which encourage the proposed model to generate more realistic and satisfying output images. To evaluate the performance of the proposed algorithm, we conducted a comprehensive study on a total of 45 sets of paired PET/CT images of clinical patients. The final experimental results demonstrated that both the generated sAC PET and sCT images showed great similarity to true AC PET and true CT images based on both qualitative and quantitative analyses. These results also indicate that in the future, our proposed algorithm has tremendous potential for reducing the need for additional anatomic imaging in hybrid PET/CT systems or the need for lengthy MR sequence acquisition in hybrid PET/MRI systems.

Design and System Evaluation of a Dual-Panel Portable PET (DP-PET)

“…Each detector module contains 5×7 blocks with 7 blocks along the y direction, while each block has 4 SiPM detector readouts coupled to an array of 7×8 15.5×2.76×2.76 mm 3 LYSO crystals. The internal light guide of the crystals is a proprietary design, and the modules were manufactured by United Imaging Healthcare Co., Ltd. (UIH), Shanghai.…”

“…The panel spacing is variable, but for this study we set it to 160 mm as this is the typical width of a pendant breast. The dimension of the FOV is 100×160×160 mm 3 , with the central FOV offset defined as (0, 0, 0) mm. Figure 1 shows the mechanical structure for one of the detector modules from the DP-PET insert.…”

“…We evaluate the interference between DP-PET and MRI from two different perspectives: the influence of the DP-PET insert on the MRI and the influence of the MRI on the DP-PET. To analyze the influence of the DP-PET on the MRI, an 8×7×12 cm 3 cuboid phantom (water, NiSO4, NaCl) was imaged using a 3D Dixon in-phase/outphase sequence (TR / TE = 4.91 / 3.19 ms, voxel size= 0.91×0.91×2 mm 3 , 549×384×164 voxels, pixel bandwith = 1080 Hz, 39s measurement time) with and without the DP-PET in the MR bore. The Signal-to-Noise Ratio (SNR) and image uniformity were assessed for the MR images acquired with and without the DP-PET in the MRI bore to determine if there was any interference caused by the DP-PET on the image quality.…”

“…Positron Emission Tomography (PET) is a biomedical imaging modality which can provide 3D molecular and functional images by detecting high energy photons generated by the positron and electron annihilation [1][2][3]. Thanks to the high sensitivity and exquisite specificity of targeted radioactive tracer, PET has been widely used in clinical settings for diagnosis of cancer, neurodegenerative disease and cardiovascular disease [4].…”

LCPR-Net: low-count PET image reconstruction using the domain transform and cycle-consistent generative adversarial networks

“…Recently, deep learning has not only been widely and successfully used in computer vision tasks but also shown great potential in the field of medical imaging (24)(25)(26)(27)(28)(29). For several different image modalities, deep learning-based reconstruction methods have been successfully applied.…”