ASCB-EMBO 2018 Poster: Deep learning enabled neurite segmentation and circuit analysis in retina dev
Updated: Jul 2, 2019
This poster will be presented by Luciano Lucas, PhD on Sunday December 9th from 12:30 - 3:00 PM. Poster number P1047.
The platform presented in this poster is a part of our latest release, Aivia Cloud. Request a demo to try it out: https://www.drvtechnologies.com/demo
The human brain is composed of approximately 100 billion neurons that collaborate to interpret our senses and control our thoughts and actions. Neurons are known to form circuits with specific functions, yet the brain’s wiring diagram remains largely unknown which greatly limits our ability to determine how degenerative diseases affect it and thus impedes the development of targeted therapies. Recently, the complete brain of the fruitfly was imaged using state of the art electron microscopy and parts of its connectome were mapped manually. However, on average each neuron took 11 hours to be traced. Tracing all 100 billion neurons in the human brain would take about 550 million years. To solve the problem, deep learning based approaches to automate neuron detection and tracing in brain samples have been proposed before. In this work we create an efficient pipeline using a 3D fully convolutional network that quickly and accurately detects neurons and traces circuits present in the mouse retina.
Our model was trained on a partially labeled image stack of size 1500x1500x75 pixels. Data augmentation was performed with random rotations by 90, 180 and 270 degrees in the XY plane and flips in X and Z dimensions. Also, each Z slice was randomly shifted along X and Y directions to simulate misalignment. Our model was trained by Adam optimizer with a base learning rate of 10-3 for 300 epochs, which took 10 hours on two NVIDIA GeForce GTX 1080Ti GPUs.
Applying the trained network to a sample image stack of 1500x1500x100 pixels (i.e. 225 million voxels) took 4.5 minutes on a single NVIDIA GeForce GTX 1080Ti GPU, thus a throughput of 50 million voxels per minute. The network generated confidence map was then processed using a seeded watershed algorithm to identify each neuron segment and the results were represented as 3D meshes. In total, we could detect 989 neuron segments in the volume tested. Neuron connections were inferred from the presence of post synaptic densities in the original image and partial neuron circuits were traced. Further editing of individual neuron boundaries or circuits is possible with arrange of semi automated tools. To greatly increase throughput, we have deployed the trained deep learning model using Aivia Cloud which currently makes uses of eight nVidia V100 GPUs. With this setup we can process approximately 4 billion voxels per minute.
Here we describe a novel machine learning enabled pipeline to greatly expedite neuron and circuit analysis. Future work includes cloud deployment of this solution to make use of multiple top performance GPUs and thus dramatically reduce the time it takes to apply the DL model to out of network data. This approach has the potential to greatly accelerate the field of connectomics.
Authors: H.Sasaki, W.Yu, C.Huang, R.Wong, J.S.Lee, L.A.Lucas
R.Wong and W.Yu are part of the Dept. of Biological Structure at the University of Washington, Seattle, WA
Other authors are a part of DRVISION Technologies LLC, Bellevue, WA