We aimed to develop a system for automatic segmentation of tumor-related vessels on ultrafast dynamic contrast (UF-DCE) MRI using U-net. Training set consisted of image dataset of 20 MIP images obtained from -15 to +60 sec after contrast injection from 59 patients. Exclusion criteria were those with poor image quality. The dice similarity coefficient was above 0.8 for training set, 0.6 for validation set. Careful analysis of failed cases revealed that inaccurate segmentation of the vessel were caused by low-contrast images, noisy-images, artifact, bright skin line due to incomplete fat suppression, and non-mass enhancement that may mimic vasculature.