Watershed learning merge tree used for segmenting neurons in SEM images

Neural circuit reconstruction is a challenging problem in neuroscience. Neuron segmentation, which recognises and segments neurons from neural images plays an important role in neural circuit reconstruction. Automated analysis method would be inevitably needed for this research because of the large amounts of image data produced by the scanning electron microscopy (SEM). In this paper, a watershed learning merge tree is used for 2D neuron segmentation. With a membrane detection map, we build a region merge tree through watershed and supervised learning algorithms, while each node of the merge tree represents a potential neuron cell. Then, a node classifier is learned to predict which node is a real neuron cell in the tree. Experimental results on automated tape-collecting ultramicrotome scanning electron microscopy (ATUM-SEM) images demonstrate the effectiveness of our method.