Objective:To reconstruct the three-dimensional (3D) dose distribution in radiotherapy based on the convolutional neural networks (CNN) through multi-perspective scintillation light processing.Methods:First, fluorescence images were captured from three orthogonal perspectives using a complementary metal-oxide-semiconductor (CMOS) imaging sensor. Then, the images were converted into 3D images, which were input to the trained CNN for dose reconstruction. Finally, the reconstructed doses in different fields were evaluated in terms of gamma pass rate, mean-square error (MSE), percentage depth dose (PDD), and cross beam profile (CBP). Additionally, as the CNN model, 3D-Unet was pre-trained on a virtual dataset.Results:With the 50% maximum dose of as the threshold and 3%/3 mm as the standard, the central-plane and stereo-mean gamma pass rates of all field reconstruction distributions were over 90%, with MSEs remained below 1%. Besides, the PDD and CBP curves showed MSEs below 1‰ and below 1%, respectively.Conclusions:The deep learning-based method for 3D dose reconstruction using scintillation light contributes to enhanced verification of instantaneous 3D relative dose based on plastic scintillation detectors.