Neural Tracing of 3D Images with Deep Learning Models
Location
PANEL: Innovative Approaches in Computer Science: Employing Technological Tools to Advance Health & Society
Science Center A126, Nancy Schrom Dye Lecture Hall
Document Type
Presentation
Start Date
4-28-2023 10:00 AM
End Date
4-28-2023 11:00 AM
Abstract
The reconstruction of 3D microscope images is vital to understand the 3D morphology of neurons and glial cells, a critical step in brain anatomy and function research. Manual and digital neural tracing has been used for investigating neurodegeneration, modulating animal behavior, and mapping out complex neural circuits. We propose a deep learning model and methods that can be further refined relative to the biological question at hand, for the analysis of morphologies, tracing, and reconstruction of neurons and glial cells. The model we propose will utilize a convolutional neural network with transformers to trace 3D images, which will then be used to extract relevant features from the images to classify different types of neurons as well as different sub-compartments of neuron and glial cells. Using datasets of annotated neuron images, Zhou and colleagues achieved average accuracy of 98% with their open software tool-box, DeepNeuron, whereas the current leading model, authored by Li and Shen only has an average accuracy of 96%. As we aim to standardize across pooled datasets and correct annotations, our model has the potential to significantly improve the efficiency and accuracy of neural classification and offer deeper insights into neural connectivity and neurodegeneration.
Keywords:
Deep learning, Neuroscience, Neural tracing, Computer science
Recommended Citation
Ates, Ada; Mar'i, Joud; and Van Valen, David, "Neural Tracing of 3D Images with Deep Learning Models" (2023). Research Symposium. 7.
https://digitalcommons.oberlin.edu/researchsymp/2023/presentations/7
Major
Neuroscience; Computer Science
Project Mentor(s)
David Van Valen and Joud Mar'I, California Institute of Technology
2023
Neural Tracing of 3D Images with Deep Learning Models
PANEL: Innovative Approaches in Computer Science: Employing Technological Tools to Advance Health & Society
Science Center A126, Nancy Schrom Dye Lecture Hall
The reconstruction of 3D microscope images is vital to understand the 3D morphology of neurons and glial cells, a critical step in brain anatomy and function research. Manual and digital neural tracing has been used for investigating neurodegeneration, modulating animal behavior, and mapping out complex neural circuits. We propose a deep learning model and methods that can be further refined relative to the biological question at hand, for the analysis of morphologies, tracing, and reconstruction of neurons and glial cells. The model we propose will utilize a convolutional neural network with transformers to trace 3D images, which will then be used to extract relevant features from the images to classify different types of neurons as well as different sub-compartments of neuron and glial cells. Using datasets of annotated neuron images, Zhou and colleagues achieved average accuracy of 98% with their open software tool-box, DeepNeuron, whereas the current leading model, authored by Li and Shen only has an average accuracy of 96%. As we aim to standardize across pooled datasets and correct annotations, our model has the potential to significantly improve the efficiency and accuracy of neural classification and offer deeper insights into neural connectivity and neurodegeneration.
