An Interactive 3D Electron Microscope Neuron Image Segmentation Framework For Large Scale Connectome Reconstruction

Reconstruction of large-scale and dense connectomes has been an essential approach towards understanding the functions of brain and revealing the mechanisms of life.However,because of the explosive scale of dataset thanks to the advances in high-resolution EM(Electron Microscopy)imaging techniques,this process is both time-consuming and labor-intensive.Specifically,recent researches point out that the primary bottleneck exists in proofreading automatically generated neuron image segmentations,which consumes tremendous resources and human efforts.Although some works have focused on improving automatic algorithms to reduce manual work,proofreading is still inevitable when considering precision and quality-control issues.Unfortunately,to the best of our knowledge,the existing interactive segmentation tools lack the ability of proofreading large segmentations in a free and scalable way,because their strategies usually have special constrains on the data,such as dividing them into relative small 3D blocks or 2D slices,and their algorithms are often not optimized.To accelerate the connectome reconstruction process,designing a more effective tool is highly demanded.In this work,we proposed a hierarchical interactive segmentation framework with optimized algorithms to tackle the challenge.To be specific,(1)the framework uses tree structure to reduce the size of neuron image progressively,extending the segmentation process from voxel-based to super-voxel-based and super-voxel-oriented,and provides corresponding low-latency and high-precision interactive segmentation algorithms as well as a time-and-space-efficient data compression algorithm.(2)For complicated segmentation situations where traditional algorithms fail to work,the framework provides an active-learning based interactive deep segmentation network to achieve high performance.(3)Last but not least,it implements a flexible computing service to manage computation-intensive tasks,which makes extending the framework with various algorithms very easily.Experimental results show that our framework has the ability to handle 100MB~1GB EM neuron images in a timely manner on normal personal computers.To conclude,this work offers an effective,scalable and expandable solution for proofreading large-scale 3D EM neuron image segmentations,which is crucial for current connectome reconstruction.