To survive and grow efficiently in harsh environments, bacterial cells make 3D colonies on surfaces, called biofilms. Biofilms are considered as the most abundant form of microbial life on Earth. To obtain an understanding of the development and heterogeneity inside biofilms, it is important to extract quantitative parameters for biofilms in space and time during growth. Single-cell quantitative analysis of images has been required to obtain the spatio-temporal analysis. During the last decade, microscopic technique significantly has improved, which enabled us to acquire 3D images of biofilm with single-cell resolution. To extract spatio-temporal information from the 3D images, individual cell inside biofilm have to be segmented. Several cell-segmentation softwares already have been used for biofilm segmentation, but those softwares still could not detect all cells in biofilms without errors. To achieve a higher accuracy of segmentation, convolutional neural network (CNN) for cell segmentation has been developed rapidly in the last few years. The CNN requires experimental raw image data and manually annotated labels as training data. The amount of accurate training data is critical for the accuracy of segmentation, but manual annotation of accurate cell labels in 3D images consumes a lot of time and manpower. Therefore, we developed the protocol to multiply accurate training dataset, to obtain 18,000 annotated cells in 3D. As a result, our trained CNN model achieved higher accuracy of segmentation than the previous software solutions. We believe that such highly accurate segmented data provides the key for single-cell tracking during biofilm growth.