A method of training neural network for obtaining medical sagittal image includes: using a first neural network on a 3-D medical image to generate a prediction sagittal mask; generating a prediction result according to the 3-D medical image and the prediction sagittal mask; generating a ground truth result according to the 3-D medical image and a ground truth sagittal mask; using a second neural network on the prediction result and the ground truth result; generating a loss function data according to an output of the second neural network; and adjusting parameters of the first neural network or the second neural network according to the loss function data.