Time-lapse or 4D seismic survey is a crucial monitoring tool for CO2 geological sequestration. Conventional time-lapse interpretation provides detailed characterization of CO2 distribution in the storage unit. However, manual interpretation is labour-intensive and often inconsistent throughout the long monitoring history, due to the inevitable changes in seismic acquisition and processing technology and interpreter’s subjectivity. We propose a neural network (NN)-based interpretation method that translates baseline and monitoring seismic images to the probability of CO2 presence. We use a simplified 3D U-Net, whose training, validation and testing are all based on the Sleipner CO2 storage project. The limited labels for training are derived from the interpreted CO2 plume outlines within the internal sandstone layers for 2010. Then we apply the trained NN on different time-lapse seismic datasets from 1999 to 2010. The results suggest that our NN-based CO2 interpretation has the following advantages: (1) high interpretation efficiency by automatic end-to-end mapping; (2) robustness against the processing-induced mismatch between the baseline and time-lapse inputs, relaxing the baseline reprocessing demands when compared to newly acquired or reprocessed time-lapse datasets; and (3) inherent interpretation consistency throughout multiple vintage datasets. Testing results with crafted time-lapse images unveil that the NN takes both amplitude difference and structural similarity into account for CO2 interpretation. We also compare 2D and 3D U-Nets under the scenario of sparse 2D labels for training. The results suggest that the 3D U-Net provides more continuous interpretation at the cost of larger computational resources for training and application.