Three-dimensional convolutional neural networks for simultaneous dual-tracer PET imaging.
Researchers
Journal
Modalities
Models
Abstract
Dual-tracer positron emission tomography (PET) is a promising technique to measure the distribution of two tracers in the body by a single scan, which can improve the clinical accuracy of disease diagnosis and can also serve as a research tool for scientists. Most current research on dual-tracer PET reconstruction is based on mixed images pre-reconstructed by algorithms, which restricts the further improvement of the precision of reconstruction. In this study, we present a hybrid loss-guided deep learning based framework for dual-tracer PET imaging using sinogram data, which can achieve reconstruction by naturally unifying two functions: the reconstruction of the mixed images and the separation of each tracer. Combined with volumetric dual-tracer images, we adopted a three-dimensional (3D) convolutional neural network (CNN) to learn full features, including spatial information and temporal information simultaneously. In addition, an auxiliary loss layer was introduced to guide the reconstruction of the dual tracers. We used Monte Carlo simulations with data augmentation to generate sucient datasets for training and testing. The results were analyzed by the bias and variance both spatially (different regions of interest) and temporally (different frames). The analysis veried the feasibility of the 3D CNN framework for dual-tracer reconstruction. Furthermore, we compared the reconstruction results with a deep belief network (DBN), which is another deep learning-based technique for the separation of dual-tracer images based on time-activity curves (TACs). The comparison results provide insights into the superior features and performance of the 3D CNN. Furthermore, we tested the [<sup>11</sup>C]FMZ-
 [<sup>11</sup>C]DTBZ images with three total-counts levels (5 x 10<sup>6</sup>, 1 x 10<sup>7</sup>, 5 x 10<sup>7</sup>), which indicate dierent noise ratios. Compared with the DBN, the 3D CNN is more robust to noise. The analysis results demonstrate that our method can successfully recover the respective distribution of lower total counts with nearly the same accuracy as that
 of the higher total counts.
© 2019 Institute of Physics and Engineering in Medicine.