An effective AI integrated system for neuron tracing on anisotropic electron microscopy volume

Abstract

Abstract Electron microscopy has become the most important technique in the field of connectomics. Several methods have been proposed in the literature to tackle the problem of dense reconstruction. However, sparse reconstruction, which is a promising technique, has not been extensively studied. As a result, we develop an AI integrated system for sparse reconstruction that can automatically trace neurons with only the initial seeded masks. First, as an important part of the system for interlayer information estimation, convolutional LSTMs are employed to estimate the spatial contexts between adjacent sections. Then, the intra-slice information is obtained by a lightweight U-Net. Moreover, we employ a novel recursive training method that can significantly improve the performance. To reduce the tracing errors caused by misalignments in large-scale data, we integrate a shift estimation and correction module that effectively improves the traced neuron length. To the best of our knowledge, this is the first attempt to apply a recurrent neural network to the task of neuron tracing. In addition, our approach performs better than other state-of-the-art methods on two highly anisotropic datasets.