Mitochondria Instance Segmentation in Electron Microscopy Image Volumes using 3D Deep Learning Networks

Segmentation of mitochondria in electron microscopy (EM) images is a challenging task due to complex shapes of mitochondria and other sub-cellular structures, background clutter, weak boundaries, low contrast, low signal-to-noise ratio, touching mitochondria, and large data size. For robust and accurate segmentation of individual mitochondria within an electron microscopy image volume, we propose a 3D deep convolutional neural network. The proposed network extends the classical U-Net semantic segmentation network with a convolutional long-short term memory (3D CLSTM U-NET). This extension allows better integration of 3D image features at different scales and abstraction levels. The proposed network generates two outputs, one corresponding to mitochondrial regions, and the other to mitochondrial boundaries. These region and boundary cues are used by a watershed segmentation module for identification of individual mitochondria. Our experiments on both animal and human datasets from MitoEM challenge showed promising results. The proposed pipeline achieved dice scores of 0.94 and 0.91 for the rat and human datasets respectively, and mAPs scores of 0.73 and 0.65.