@article {Wen385567, author = {Chentao Wen and Takuya Miura and Yukako Fujie and Takayuki Teramoto and Takeshi Ishihara and Koutarou D. Kimura}, title = {Deep-learning-based flexible pipeline for segmenting and tracking cells in 3D image time series for whole brain imaging}, elocation-id = {385567}, year = {2018}, doi = {10.1101/385567}, publisher = {Cold Spring Harbor Laboratory}, abstract = {The brain is a complex system that operates based on coordinated neuronal activities. Brain-wide cellular calcium imaging techniques have quickly advanced in recent years and become powerful tools for understanding the neuronal activities of small animal models. The whole brain imaging generally requires to extract the neuronal activities from three-dimensional (3D) image series. Unfortunately, the 3D image series are obtained under imaging conditions different among laboratories and extracting neuronal activities from the data requires multiple processes. Therefore researchers need to develop their own software, which has prevented the application of whole-brain imaging experiments in more laboratories. Here, we combined traditional image processing techniques with the powerful deep-learning method which can be flexibly modified to fit 3D image data in the nematode Caenorhabditis elegans obtained under different conditions. We first trained the 3D U-net deep network to classify each pixel into cell and non-cell categories. Cells merged as a whole region were further separated into individual cells by watershed segmentation. The cells were then tracked in 3D space over time with the combination of a feedforward network and a point set registration method to use local and global relative positions of the cells, respectively. Remarkably, one manually annotated 3D image combined with data augmentation was sufficient for training the deep networks to obtain satisfactory tracking results. Our method correctly tracked more than 98\% of neurons in three different image datasets and successfully extracted brain-wide neuronal activities. Our method worked well even when the sampling rate was reduced: 86\% correct in case 4/5 frames were removed, and when artificial noise was added into the raw images: 91\% correct in case 35 times of background-level noise was added. Our results proved that deep learning is widely applicable to different datasets and can help us in establishing a flexible pipeline for extracting whole brain activities.}, URL = {https://www.biorxiv.org/content/early/2018/08/06/385567}, eprint = {https://www.biorxiv.org/content/early/2018/08/06/385567.full.pdf}, journal = {bioRxiv} }