BIOMEDICAL SEGMENTATION ON CELL AND BRAIN IMAGES

The biomedical imaging techniques grow rapidly and output big amount of data quickly in the recent years. But image segmentation, one of the most important and fundamental biomedical data analysis techniques, is still time-consuming for human annotators. Therefore, there is an urgent need for segmentation to be taken by machine automatically. Segmentation is essential for biomedical image analysis and could help researchers to gain further diagnostic insights. This paper has three topics under biomedical image segmentation scenario. For the first topic, we examine a popular deep learning structure for segmentation task, U-Net, and modify it for our task on bacteria cell images by using boundary label setting and weighted loss function. Compared to the MATLAB segmentation program used before, the new deep learning method improves the performance in terms of object-level evaluation metrics. For the second topic, we participate into a brain image segmentation challenge which aims for helping neuroscientists to segment the membrane from neurites in order to get the reconstruction of neurites circuit. Data augmentation tricks and multiple loss functions are examined for improving the test performance and finally using combined loss functions can out-perform the original U-Net result in terms of the official ranking metric. A new dice loss is designed to focus more on the hard to segment class. The third topic is to apply the unsupervised segmentation method which will not be restrained by human labelling speed and effort. This is meaningful under biomedical segmentation scenario where training data with expert labelling is always lacking. Without using any labelled data, the unsupervised method, Double DIP, performs better than the MATLAB program on the semantic level