We have developed a new dosimetry approach, called CFDose, for liver cancer radioembolization based on computational fluid dynamics (CFD) simulation in the hepatic arterial tree. Although CFDose overcomes some of the limitations of the current dosimetry methods such as the unrealistic assumption of homogeneous distribution of yttrium-90 in the liver, it suffers from the expensive computational cost of CFD simulations. To accelerate CFDose, we introduce a deep learning model to predict the blood flow distribution between the liver segments in a patient with hepatocellular carcinoma. The model was trained with the results of CFD simulations under different outlet boundary conditions. The model consisted of convolutional, average pooling and transposed convolution layers. A regression layer with a mean-squared-error loss function was utilized at the network output to estimate the arterial outlet blood flow. The mean-squared error and prediction accuracy were calculated to measure model performance. Results showed that the average difference between the CFD results and predicted flow data was less than 2.45% for all the samples in the test dataset. The proposed model thus estimated the blood flow distribution with high accuracy significantly faster than a CFD simulation. The network output can be used to estimate the yttrium-90 dose distribution in the liver in future studies.