A system-level DOI discrimination method for PET imaging based on semi-supervised learning

We propose a system-level DOI discrimination method based on semi-supervised learning, for developing a TOF-DOI human brain dedicated PET system using a multi-resolution detector. Existing DOI discrimination methods require the acquisition of side-irradiation data from each detector, which was difficult to achieve in the PET system. To obtain a pre-training CNN model for DOI discrimination, the top and bottom layer LYSO events was acquired from a base-detector using electronically collimated side-irradiation, with the corresponding data labeled according to ground truth DOI layers. Then, using regular top-irradiation data of the test-detector (the detector we intend to generate a model for) and the side-irradiation data of the base-detector, the pre-trained model was further trained with a semi-supervised learning algorithm to produce an adapted CNN model suitable for the test-detector. The method’s reliability and robustness were tested with multiple detectors from different manufacturers, as well as by changing the acquisition system parameters to mimic the variabilities between detectors. The method proposed was all able to accurately discriminate between top and bottom LYSO events for the top-irradiation data of the detector. This method significantly reduces the data acquisition time and the labeling cost associated with deep learning for system-level DOI discrimination: Simply by using top-irradiation data from all test-detectors and side-irradiation data from one (or a much smaller number) base-detector, DOI discrimination models can be acquired for each detector in a PET system. By used this technique, DOI discrimination methods based on deep learning are made more effective and practical.