Artificial Intelligence-Based Automated Neural Cell Phenotype Identification
Abstract
Approved for Public ReleaseTraumatic brain injury (TBI) from directed high-power pulsed radiofrequency/microwave energy exposure poses a serious threat to civilian and military populations. Among the several TBI types, including mild TBI (e.g., concussion) and blast-related TBI, directed energy (DE)-related TBI is the least understood, partly due to the intricate nature of exposure conditionsand potential electro-mechanical pathways to injury at the cellular scale. Understanding DE-related injury thresholds at the cellular level is a critical and currently missing component for understanding this condition, developing accurate injury prediction, and designing tools and approaches for adequate protection from this injury. Cellular-scale TBI studies can fill this gap by analyzing the behavior of individual neural cell phenotypes in 3D polyculture systems when exposed to RF or microwave energy at different powerdensities and loading frequencies for various time durations. Such studies are not readily feasible with currently available labor-intensive manual immunolabeling, imaging, segmentation, and post-processing steps. The primary objective of the proposed project is to overcome this roadblock by developing a framework for rapidly identifying different neural phenotypes from 3D confocal micrographs based on a single, non-specific cell marker.In the first stage of this project, we will develop a machine learning (ML)-based model for neural cell phenotype identification and segmentation (Aim 1). Preliminary training data for these models will be generated via confocal microscopy of 3D neural polycultures labeled with (i) calcein AM (input data, showing all living cells without differentiating), and (ii) seven different neural cell phenotype-specific markers (target data, showing specific phenotypes separately). Giventhe challenges associated with limited training data, this data will be augmented to create a larger dataset appropriate for final model training. Two supervised learning frameworks, an ML model combining nonlinear dimensionality reduction and Gaussian Process Regression (GPR) and a deep learning model combining 3D U-Net and conditional generative adversarial network (cGAN), will be implemented in parallel, and their performances will be compared. Overall, Aim 1 work will generate ML models (one for each cell phenotype) that can identify (i.e., label and segment) specific neural cell phenotypes from a single confocal micrograph of 3D neural polyculture stained with calcein AM.In the second stage, we will create an open-source Python package that houses an ensemble of ML models trained in Aim 1 work for cell-type identification and offers useful functionalities for injury studies, such as analyses of cell morphology and motion/deformation. Specifically, the ensemble model will take a single 3D calcein AM image and provide a merged image output, showing different neural cell phenotypes.Upon successfully completing this research project, we will have published the first automatic neural cell phenotype identification and analysis tool for cellular-scale TBI research. This tool and the insights generated from its development will create a long-term impact on the field by accelerating research on the understanding of phenotype-specific biomechanical behavior and injury thresholds for DE-related TBI. An understanding of the intrinsic cell type-specific injury thresholds is key to quantifying brain location-dependent vulnerabilities and designing the next generation of head protection devices for the warfighter.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 09, 2024
- Source ID
- N000142412790
Entities
People
- Kshitiz Upadhyay
Organizations
- Louisiana State University
- Office of Naval Research
- United States Navy