Recursive Training of 2D-3D Convolutional Networks for Neuronal Boundary\n Detection

2015/08/19 by Kisuk Lee, Lee, Kisuk, Aleksandar Zlateski +5 · 1 citation
Biochemistry, Genetics and Molecular Biology · Computer Science · Materials Science · Mathematics · #2D Filters #Advanced Electron Microscopy Techniques and Applications #Advanced Neural Network Applications #Algorithm #Artificial intelligence #Artificial neural network #Backpropagation #Boundary (topology) #Cell Image Analysis Techniques #Computation #Computer Vision and Pattern Recognition (cs.CV) #Computer science #Computer vision #Connectome #Connectomics #Context (archaeology) #Convolutional neural network #Deep learning #Electron and X-Ray Spectroscopy Techniques #FOS: Computer and information sciences #Image segmentation #Mathematics #Network architecture #Object detection #Parallel computing #Pattern recognition (psychology) #Pooling #Segmentation #cs.CV

paper · pdf · doi:10.48550/arxiv.1508.04843

published in arXiv (Cornell University) 28, 3573-3581 (Cornell University)

openalex publication_date 2015/08/19 · arxiv created 2015/08/20 · arxiv updated 2015/08/21 · openalex created_date 2025/10/10 · openalex updated_date 2026/08/06

Abstract

Efforts to automate the reconstruction of neural circuits from 3D electron\nmicroscopic (EM) brain images are critical for the field of connectomics. An\nimportant computation for reconstruction is the detection of neuronal\nboundaries. Images acquired by serial section EM, a leading 3D EM technique,\nare highly anisotropic, with inferior quality along the third dimension. For\nsuch images, the 2D max-pooling convolutional network has set the standard for\nperformance at boundary detection. Here we achieve a substantial gain in\naccuracy through three innovations. Following the trend towards deeper networks\nfor object recognition, we use a much deeper network than previously employed\nfor boundary detection. Second, we incorporate 3D as well as 2D filters, to\nenable computations that use 3D context. Finally, we adopt a recursively\ntrained architecture in which a first network generates a preliminary boundary\nmap that is provided as input along with the original image to a second network\nthat generates a final boundary map. Backpropagation training is accelerated by\nZNN, a new implementation of 3D convolutional networks that uses multicore CPU\nparallelism for speed. Our hybrid 2D-3D architecture could be more generally\napplicable to other types of anisotropic 3D images, including video, and our\nrecursive framework for any image labeling problem.\n

Citations

• Saturated Reconstruction of a Volume of Neocortex
• Maximin affinity learning of image segmentation
• Connectomic reconstruction of the inner plexiform layer in the mouse retina
• ImageNet classification with deep convolutional neural networks
• Learning Spatiotemporal Features with 3D Convolutional Networks

Cited by

• A Review of Network Inference Techniques for Neural Activation Time\n Series

Related

• Multi-stage Multi-recursive-input Fully Convolutional Networks for Neuronal Boundary Detection
• Deep and Wide Multiscale Recursive Networks for Robust Image Labeling
• Convolutional nets for reconstructing neural circuits from brain images acquired by serial section electron microscopy
• Learning and Segmenting Dense Voxel Embeddings for 3D Neuron Reconstruction
• An Error Detection and Correction Framework for Connectomics