StemNet : A Temporally Trained Fully Convolutional Network for Segmentation of Muscular Stem Cells

Sammanfattning: In biomedical research, time-lapse microscopy is an important tool to be able tostudy processes which are too slow for humans to observe. This technique ispowerful since it gives information about how parameters of single cells changeover time.The problem to be solved in this project is to segment MuSCs (Muscular Stem Cells)in images and to classify them. This is done by using a deep neural network trainedusing supervised learning. The network is inspired by the architecture of the U-net,but extended by using temporal data to see if it can increase its performance. Thenetwork is trained on images from the time-lapse sequence, where the temporalaspect is used to create a short-term memory for the network. The results arecompared to a network of the same architecture but without the temporal aspect inthe training.The temporal approach shows that the network learns faster what is roughly aMuSC and what is not, but in the end it gives a slightly higher and more accurateclassification of MuSCs by training the network without giving it a short-termmemory, for this task.