Multiphase fluid flow in porous media has been extensively studied for its applications in carbon capture and storage, hydrocarbon recovery, aquifer contamination, soil hydrology and subsurface energy resources. Fluid displacement in porous media can be investigated using highly time-resolved synchrotron X-ray microtomography. One consequence of extremely fast imaging can be compromised image quality, including noise and decreased contrast, which makes the images hard to be segmented. We trained an established convolutional neural network (CNN) architecture (U-Net) with 18,072 images from multiphase flow experiments generated by synchrotron μCT. The trained neural network can segment synchrotron μCT images of core-flooding experiments rapidly and accurately without any pre-processing of the raw image. Segmenting one μCT scan volume of size 1004x496x496 takes 5.6 minutes on an Nvidia Quadro K5200 GPU, while a conventional segmentation pipeline using CPU for the same size data takes 50.2 minutes. On the test dataset, the AUC-ROC score of individual class reached above 0.99 and the mean accuracy of the three segmentation classes reached 99%. The average IoU of the three classes is 0.98. The accuracy of the CNN segmentation is of the same order as conventional methods but it is significantly faster.