Boron concentration prediction from Compton camera image for boron neutron capture therapy based on generative adversarial network

Abstract

Prompt gamma monitoring for the prediction of boron concentration is valuable for the dose calculation of boron neutron capture therapy (BNCT). This work proposes to use generative adversarial network (GAN) to predict the boron distribution based on Compton camera (CC) imaging quickly and provide a scientific basis for its application in BNCT. The BNCT and Compton imaging process was simulated, then the image reconstructed from the simulation and the contour of skin from CT are used as input, and the distribution of boron concentration from PET data is set as the output to train the network. The structural similarity, peak signal-to-noise ratio, and root mean square error of the images generated by the trained network are improved significantly, and the ratio of the boron concentration between the tumor area and the normal tissue is improved from 1.55 to 3.85, which is much closer to the true value of 3.52. The trained network can optimize the original image within 0.83 s, which is much faster than iterative optimization. The proposed method could help to ease the current online monitoring problem of boron concentration on a computational level, thereby promoting the clinical development of BNCT technology.