Deep learning-based image reconstruction for TOF PET with DIRECT data partitioning format

Conventional positron emission tomography (PET) image reconstruction is achieved by the statistical iterative method. Deep learning provides another opportunity for speeding up the image reconstruction process. However, conventional deep learning-based image reconstruction requires a fully connected network for learning the Radon transform. The use of fully connected networks greatly complicated the network and increased hardware cost. In this study, we proposed a novel deep learning-based image reconstruction method by utilizing the DIRECT data partitioning method. The U-net structure with only convolutional layers was used in our approach. Patch-based model training and testing were used to achieve 3D reconstructions within current hardware limitations. Time-of-flight (TOF)-histoimages were first generated from the listmode data to replace conventional sinograms. Different projection angles were used as different channels in the input. A total of 15 patient data were used in this study. For each patient, the dynamic whole-body scanning protocol was used to expand the training dataset and a total of 372 separate scans were included. The leave-one-patient-out validation method was used. Two separate studies were carried out. In the first study, the measured TOF-histoimages were directly used for model training and testing, to study the performance of the method in real-world applications. In the second study, TOF-histoimages were simulated from already reconstructed images to exclude the scatters, randoms, attenuation-activity mismatch effects. This study was used to evaluate the optimal performance when all other corrections are ideal. Volumes of interests were placed in the liver and lesion region to study image noise and lesion quantitations. The reconstructed images using the proposed deep learning method showed similar image quality when compared with the conventional expectation-maximization approach. A minimal difference was observed when the simulated TOF-histoimages were used as model input and testing, suggesting the deep learning model can indeed learn the reconstruction process. Some quantitative difference was observed when the measured TOF-histoimages were used. The two studies suggested that the major difference is caused by inaccurate corrections performed by the network itself, which indicated that physics-based corrections are still required for better quantitative performance. In conclusion, we have proposed a novel deep learning-based image reconstruction method for TOF PET. With the help of the DIRECT data partitioning method, no fully connected layers were used and 3D image reconstruction can be directly achieved within the limits of the current hardware.