b-Net: a 3D neural network for unsupervised instance segmentation of brainbow

Abstract

Image segmentation is one of the most important tasks in medical image analysis and is the starting point to understand how brain functions. Recently studies have been focused on supervised deep segmentation algorithms, but in the brain image domain, sufficient pixel-level label data are hard to obtain. In this paper, focusing on the color information of Brainbow data where each neurons are stained into different colors, and address the problem with unsupervised instance segmentation, presenting b-Net, a more powerful architecture for Brainbow image instance segmentation. Our architecture is a 3D encoder-decoder network where it properly function as auto-encoder through sparsity constraint, where encoder produces a k-way pixel-wise prediction, providing meaningful segmentation results in the ProExM Brainbow neuron segmentation. Here encoder is also trained to function as CRF, utilizing the label consistency for segmentation. It is shown that the proposed model can provide proper segmentation for each neuron.