Abstract Methods Using Deep Neural Networks to Understand Neural ...
Graduate Category: Physical and Life Sciences Degree Seeking: PhD Abstract ID #1867
Using Deep Neural Networks to Understand Neural Networks in the Brain Shih-Luen Wang, Rohan Gala, Seyed Mostafa Mousavi Kahaki, and Armen Stepanyants Department of Physics and Center for Interdisciplinary Research on Complex Systems, Northeastern University, Boston, MA 02115
Abstract Unraveling the incredibly complex wiring diagram of billions of neurons that are interconnected by trillions of synapses is an overwhelming challenge that must be overcome if we are to understand how brain structure is related to function. Recent advances in genetic engineering and optical microscopy allow neuroscientists to label and image a sparse subset of neurons at sub-micron resolution in 3-D over the entire brain volume. Such imaging experiments produce terabytes of data, and at present, manually tracing the neuronal arbors using software tools is the only reliable way to extract quantitative information from raw images. These manual tracing efforts are extremely time consuming and are prone to user bias, and therefore, they are unsuitable for high throughput experiments. Here, we present an algorithm to automatically reconstruct neuronal arbors from 3-D image stacks. Our method involves artificial neural networks to learn features that characterize neuronal arbors and thereby enhance noisy images. This step is followed by a customized path finding algorithm to recover the underlying neuronal morphology. We show that our automated approach produces results that are comparable to manual traces in terms of accuracy and provide tremendous improvement in tracing time.
Methods
Raw image
segmentation
Results
Trace finding
Topology finding
We break down the reconstruction problem into 3 simpler steps. Segmentation aim to separate pixels that belongs to the foreground from the background. Trace finding aim to connect the centerline of foregrounds. Topology finding aim to figure out how the traces are connected to each others. We apply artificial neural network (deep learning) to solve segmentation problem
Our deep learning approach on segmentation can find very dim axons and clean off the noises.
Problem statement Artificial neuron Decision boundary
Class 1
g
O g WV i i i
New decision boundary Class 1
Old decision boundary
O V2
V2 Class 2
Class 2
Wi Vi
Reconstructing neuron connectivity is essential for understanding the brain There are around 100 billion neuron cells in an adult brain, around 100 million in a mice brain. 1mm3 brain tissue contains nearly 105 neurons along with 4.5 km of axons and dendrites. Nowadays, experimentalist can image a whole rodent brain in a day and generate nearly 20 Terabytes of data.
V1
V1
An artificial neuron is a mathematical function that models the biological neuron. Though training, the artificial neuron will adjust its decision boundary (Wi) to maximally classify training examples we give. This is a supervised learning strategy. We use sigmoid function as the activation function of our artificial neuron.
0.2
0.1
0
-0.1 -0.2
… …
1 Neuron (output)
21
Neuron connectivity is highly non-local. Trace a single neuron need to follow the trace of an axon or dendrite over hundreds of image stacks, a mistake in any stack can lead to incorrect reconstruction result. Neuron tracing is a 3D image procession problem.
21 learned weights
Our result is even more accurate than human tracers. Our deep learning approach is doing way better simple thresholding, which will be very useful for trace finding.
Ongoing Work Base on the segmentation result, we plan to apply path finding algorithm such as A* or Dijkstra algorithms to find the trace We are collaborating with computer scientists to make a user-friendly software for neuron tracing, and plan to include our automated tracing program in it.
Impact
21
Intensity image (input)
Our result User 1 User 2 Thresholding
36 Neurons (hidden layer)
We want our network to learn weather a pixel is a foreground or not by input a 3D sub-image of 21 by 21 by 7 centered at that pixel. We extracted around 107 such sub-images, separate them into not intersecting training and test data sets. We trained the network for around 10 hours and get the learned weights. After training, the network can segment more than 104 examples in a second.
Our automatic neuron tracing program aim to save a huge amount of human labor from manual tracing task and generate large datasets of the brain connectivity. A larger dataset of the brain connectivity can help us learn more about how our brain works, treat disease like Alzheimer's and build better computer devices.
References Shepherd et al., Nature Neuroscience 2005 Gala et al., Frontiers in Neuroanatomy 2014 Economo et al., eLife 2016
Additional information The Neural Circuit Tracer software package is available at www.neurogeometry.net. This work was supported by NIH grant R01 NS091421.
Using Deep Neural Networks to Understand Neural Networks in the Brain Shih-Luen Wang, Rohan Gala, Seyed Mostafa Mousavi Kahaki, and Armen Stepanyants Department of Physics and Center for Interdisciplinary Research on Complex Systems, Northeastern University, Boston, MA 02115
Abstract Unraveling the incredibly complex wiring diagram of billions of neurons that are interconnected by trillions of synapses is an overwhelming challenge that must be overcome if we are to understand how brain structure is related to function. Recent advances in genetic engineering and optical microscopy allow neuroscientists to label and image a sparse subset of neurons at sub-micron resolution in 3-D over the entire brain volume. Such imaging experiments produce terabytes of data, and at present, manually tracing the neuronal arbors using software tools is the only reliable way to extract quantitative information from raw images. These manual tracing efforts are extremely time consuming and are prone to user bias, and therefore, they are unsuitable for high throughput experiments. Here, we present an algorithm to automatically reconstruct neuronal arbors from 3-D image stacks. Our method involves artificial neural networks to learn features that characterize neuronal arbors and thereby enhance noisy images. This step is followed by a customized path finding algorithm to recover the underlying neuronal morphology. We show that our automated approach produces results that are comparable to manual traces in terms of accuracy and provide tremendous improvement in tracing time.
Methods
Raw image
segmentation
Results
Trace finding
Topology finding
We break down the reconstruction problem into 3 simpler steps. Segmentation aim to separate pixels that belongs to the foreground from the background. Trace finding aim to connect the centerline of foregrounds. Topology finding aim to figure out how the traces are connected to each others. We apply artificial neural network (deep learning) to solve segmentation problem
Our deep learning approach on segmentation can find very dim axons and clean off the noises.
Problem statement Artificial neuron Decision boundary
Class 1
g
O g WV i i i
New decision boundary Class 1
Old decision boundary
O V2
V2 Class 2
Class 2
Wi Vi
Reconstructing neuron connectivity is essential for understanding the brain There are around 100 billion neuron cells in an adult brain, around 100 million in a mice brain. 1mm3 brain tissue contains nearly 105 neurons along with 4.5 km of axons and dendrites. Nowadays, experimentalist can image a whole rodent brain in a day and generate nearly 20 Terabytes of data.
V1
V1
An artificial neuron is a mathematical function that models the biological neuron. Though training, the artificial neuron will adjust its decision boundary (Wi) to maximally classify training examples we give. This is a supervised learning strategy. We use sigmoid function as the activation function of our artificial neuron.
0.2
0.1
0
-0.1 -0.2
… …
1 Neuron (output)
21
Neuron connectivity is highly non-local. Trace a single neuron need to follow the trace of an axon or dendrite over hundreds of image stacks, a mistake in any stack can lead to incorrect reconstruction result. Neuron tracing is a 3D image procession problem.
21 learned weights
Our result is even more accurate than human tracers. Our deep learning approach is doing way better simple thresholding, which will be very useful for trace finding.
Ongoing Work Base on the segmentation result, we plan to apply path finding algorithm such as A* or Dijkstra algorithms to find the trace We are collaborating with computer scientists to make a user-friendly software for neuron tracing, and plan to include our automated tracing program in it.
Impact
21
Intensity image (input)
Our result User 1 User 2 Thresholding
36 Neurons (hidden layer)
We want our network to learn weather a pixel is a foreground or not by input a 3D sub-image of 21 by 21 by 7 centered at that pixel. We extracted around 107 such sub-images, separate them into not intersecting training and test data sets. We trained the network for around 10 hours and get the learned weights. After training, the network can segment more than 104 examples in a second.
Our automatic neuron tracing program aim to save a huge amount of human labor from manual tracing task and generate large datasets of the brain connectivity. A larger dataset of the brain connectivity can help us learn more about how our brain works, treat disease like Alzheimer's and build better computer devices.
References Shepherd et al., Nature Neuroscience 2005 Gala et al., Frontiers in Neuroanatomy 2014 Economo et al., eLife 2016
Additional information The Neural Circuit Tracer software package is available at www.neurogeometry.net. This work was supported by NIH grant R01 NS091421.
