Feasibility study on image reconstruction for single-photon emission computed tomography with limited projections by neural networks

Abstract

This study investigated the feasibility of using a neural network for single-photon emission computed tomography (SPECT) in boron neutron capture therapy (BNCT) with limited projection data. In order to determine the therapeutic dose for use in BNCT under neutron irradiation, a real-time reconstruction method is required. However, BNCT-SPECT has limitations related to the projection angle range and number of projection points. It is not easy to reconstruct an image under these projection limitations using a conventional filtered back-projection (FBP) method. Therefore, a convolutional neural network (CNN) was adopted to reconstruct radioactivity distribution images using the responses from radiation detectors arranged to surround a system. An original dataset of radioactivity distribution images with specific conditions was prepared, and the CNN was trained using the training data. The performances of the CNN for BNCT-SPECT with the limited projections were evaluated. The conditions were the limitations on the number of projections in 180° and on the projection angle range at a constant step angle. The results were compared with those of the FBP method and maximum likelihood-expectation maximization (ML-EM) method. The CNN could be used to estimate radioactivity distributions for BNCT-SPECT with a limited projection angle range. Under the condition of only five projection angles in 180°, the CNN and ML-EM methods could reconstruct images more accurately than the FBP method. In the case of the limitation on the projection angle range, the tolerance of the projection angle range with a step angle of 5° for CNNs was greater than approximately 45°.