Supervoxel-Based Segmentation of Mitochondria in EM Image Stacks With Learned Shape Features

摘要: It is becoming increasingly clear that mitochondria play an important role in neural function. Recent studies show mitochondrial morphology to be crucial cellular physiology and synaptic function a link between defects neuro-degenerative diseases strongly suspected. Electron microscopy (EM), with its very high resolution all three directions, one of the key tools look more closely into these issues but huge amounts data it produces make automated analysis necessary. State-of-the-art computer vision algorithms designed operate on natural 2-D images tend perform poorly when applied EM for number reasons. First, sheer size typical volume renders most modern segmentation schemes intractable. Furthermore, approaches ignore shape cues, relying only local statistics easily become confused confronted noise textures inherent data. Finally, conventional assumption strong image gradients always correspond object boundaries violated by clutter distracting membranes. In this work, we propose graph partitioning scheme addresses issues. reduces computational complexity operating supervoxels instead voxels, incorporates features capable describing 3-D target objects, learns recognize distinctive appearance true boundaries. Our experiments demonstrate our approach able segment at performance level close human annotator, outperforms state-of-the-art technique.