SEARCH · Search NASA

Results for “overfitting”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

SourceResearch connectionFrom yearThrough year

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 109 records · Page 6

All sources NASA NTRS DOE OSTI

DOE OSTI2022

Towards a robust out-of-the-box neural network model for genomic data

The accurate prediction of biological features from genomic data is paramount for precision medicine and sustainable agriculture. For decades, neural network models have been widely popular in fields like computer vision, astrophysics and targeted marketing given their prediction accuracy and their robust performance under big data settings. Yet neural network models have not made a successful transition into the medical and biological world due to the ubiquitous characteristics of biological data such as modest sample sizes, sparsity, and extreme heterogeneity. Here, we investigate the robustness, generalization potential and prediction accuracy of widely used convolutional neural network and natural language processing models with a variety of heterogeneous genomic datasets. Mainly, recurrent neural network models outperform convolutional neural network models in terms of prediction accuracy, overfitting and transferability across the datasets under study. While the perspective of a robust out-of-the-box neural network model is out of reach, we identify certain model characteristics that translate well across datasets and could serve as a baseline model for translational researchers.

59 BASIC BIOLOGICAL SCIENCES↗

DOE OSTI2025

Evaluation of the 2022 West Nile virus forecasting challenge, USA

Abstract Background West Nile virus (WNV) is the most common cause of mosquito-borne disease in the continental USA, with an average of ~1200 severe, neuroinvasive cases reported annually from 2005 to 2021 (range 386–2873). Despite this burden, efforts to forecast WNV disease to inform public health measures to reduce disease incidence have had limited success. Here, we analyze forecasts submitted to the 2022 WNV Forecasting Challenge, a follow-up to the 2020 WNV Forecasting Challenge. Methods Forecasting teams submitted probabilistic forecasts of annual West Nile virus neuroinvasive disease (WNND) cases for each county in the continental USA for the 2022 WNV season. We assessed the skill of team-specific forecasts, baseline forecasts, and an ensemble created from team-specific forecasts. We then characterized the impact of model characteristics and county-specific contextual factors (e.g., population) on forecast skill. Results Ensemble forecasts for 2022 anticipated a season at or below median long-term WNND incidence for nearly all (> 99%) counties. More counties reported higher case numbers than anticipated by the ensemble forecast median, but national caseload (826) was well below the 10-year median (1386). Forecast skill was highest for the ensemble forecast, though the historical negative binomial baseline model and several team-submitted forecasts had similar forecast skill. Forecasts utilizing regression-based frameworks tended to have more skill than those that did not and models using climate, mosquito surveillance, demographic, or avian data had less skill than those that did not, potentially due to overfitting. County-contextual analysis showed strong relationships with the number of years that WNND had been reported and permutation entropy (historical variability). Evaluations based on weighted interval score and logarithmic scoring metrics produced similar results. Conclusions The relative success of the ensemble forecast, the best forecast for 2022, suggests potential gains in community ability to forecast WNV, an improvement from the 2020 Challenge. Similar to the previous challenge, however, our results indicate that skill was still limited with general underprediction despite a relative low incidence year. Potential opportunities for improvement include refining mechanistic approaches, integrating additional data sources, and considering different approaches for areas with and without previous cases. Graphical Abstract

54 ENVIRONMENTAL SCIENCES↗

DOE OSTI2026

Learning genetic perturbation effects with variational causal inference

Advances in sequencing technologies have enhanced the understanding of gene regulation in cells. In particular, Perturb-seq has enabled high-resolution profiling of the transcriptomic response to genetic perturbations at the single-cell level. This understanding has implications in functional genomics and potentially for identifying therapeutic targets. Various computational models have been developed to predict perturbational effects. While deep learning models excel at interpolating observed perturbational data, they tend to overfit in the lack of enough data and may not generalize well to unseen perturbations. In contrast, mechanistic models, such as linear causal models based on gene regulatory networks, hold greater potential for extrapolation, as they encapsulate regulatory information that can predict responses to unseen perturbations. However, their application has been limited to small studies due to overly simplistic assumptions, making them less effective in handling noisy, large-scale single-cell data. We propose a hybrid approach that combines a mechanistic causal model with variational deep learning, termed Single Cell Causal Variational Autoencoder (SCCVAE). The mechanistic model employs a learned regulatory network to represent perturbational changes as shift interventions that propagate through the learned network. SCCVAE integrates this mechanistic causal model into a variational autoencoder, generating rich, comprehensive transcriptomic responses. Our results indicate that SCCVAE exhibits superior performance over current state-of-the-art baselines for extrapolating to predict unseen perturbational responses. Additionally, for the observed perturbations, the latent space learned by SCCVAE allows for the identification of functional perturbation modules and simulation of single-gene knockdown experiments of varying penetrance, presenting a robust tool for interpreting and interpolating perturbational responses at the single-cell level.

59 BASIC BIOLOGICAL SCIENCES↗

DOE OSTI2020

BASS: An R Package for Fitting and Performing Sensitivity Analysis of Bayesian Adaptive Spline Surfaces

In this work, we present the R package BASS as a tool for nonparametric regression. The primary focus of the package is fitting fully Bayesian adaptive spline surface (BASS) models and performing global sensitivity analyses of these models. The BASS framework is similar to that of Bayesian multivariate adaptive regression splines (BMARS) from Denison, Mallick, and Smith (1998), but with many added features. The software is built to efficiently handle significant amounts of data with many continuous or categorical predictors and with functional response. Under our Bayesian framework, most priors are automatic but these can be modified by the user to focus on parsimony and the avoidance of overfitting. If directed to do so, the software uses parallel tempering to improve the reversible jump Markov chain Monte Carlo (RJMCMC) methods used to perform inference. We discuss the implementation of these features and present the performance of BASS in a number of analyses of simulated and real data.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2020

TBASS: A Robust Adaptation of Bayesian Adaptive Spline Surfaces

The R package TBASS is an extension of the BASS package created by Francom and Sansó (2019). The package is used to fit a Bayesian multivariate adaptive spline to a dataset that either follows a Student’s t-distribution or has outliers. Much of the framework for TBASS is adapted from the concepts of Bayesian Multivariate Adaptive Regression Splines (BMARS), specifically the work done by Denison, Mallick, and Smith (1998). The spline function is fit using a Reversible-Jump Markov Chain Monte Carlo algorithm,. By including this more robust generalization, a dataset with outliers can be accurately fit using the BMARS model, without the possibility of overfitting or variance inflation.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2020

Limited-view Cone Beam CT reconstruction using 3D Patch-based Supervised and Adversarial Learning [Slides]

We present a novel machine learning CNN architecture that can learn from limited data combined appropriately with physics and statistical priors (e.g., forward models and noise models). To address the limited availability of training data we adopt a 3D patch-based approach for our models. Patch-based learning is central to several image reconstruction methods and demands fewer training data than DL approaches, as a single data volume can be broken into several millions of overlapping 3D sub-volumes or patches. This creates a very large number of training sub-volumes from a limited number of overall image volumes. A 3D Generative Adversarial Networks (GAN) is then trained to remove artifacts at the sub-volume level. The combination of a sub-volume-based approach with DL allows us to exploit the richness of the latter in extracting and representing image features, while avoiding risks associated with overfitting due to limited training data.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

DOE OSTI2022

Computational Imaging for Intelligence in Highly Scattering Aerosols (Final Report)

Natural and man-made degraded visual environments pose major threats to national security. The random scattering and absorption of light by tiny particles suspended in the air reduces situational awareness and causes unacceptable down-time for critical systems and operations. To improve the situation, we have developed several approaches to interpret the information contained within scattered light to enhance sensing and imaging in scattering media. These approaches were tested at the Sandia National Laboratory Fog Chamber facility and with tabletop fog chambers. Computationally efficient light transport models were developed and leveraged for computational sensing. The models are based on a weak angular dependence approximation to the Boltzmann or radiative transfer equation that appears to be applicable in both the moderate and highly scattering regimes. After the new model was experimentally validated, statistical approaches for detection, localization, and imaging of objects hidden in fog were developed and demonstrated. A binary hypothesis test and the Neyman-Pearson lemma provided the highest theoretically possible probability of detection for a specified false alarm rate and signal-to-noise ratio. Maximum likelihood estimation allowed estimation of the fog optical properties as well as the position, size, and reflection coefficient of an object in fog. A computational dehazing approach was implemented to reduce the effects of scatter on images, making object features more readily discernible. We have developed, characterized, and deployed a new Tabletop Fog Chamber capable of repeatably generating multiple unique fog-analogues for optical testing in degraded visual environments. We characterized this chamber using both optical and microphysical techniques. In doing so we have explored the ability of droplet nucleation theory to describe the aerosols generated within the chamber, as well as Mie scattering theory to describe the attenuation of light by said aerosols, and correlated the aerosol microphysics to optical properties such as transmission and meteorological optical range (MOR). This chamber has proved highly valuable and has supported multiple efforts inclusive to and exclusive of this LDRD project to test optics in degraded visual environments. Circularly polarized light has been found to maintain its polarization state better than linearly polarized light when propagating through fog. This was demonstrated experimentally in both the visible and short-wave infrared (SWIR) by imaging targets made of different commercially available retroreflective films. It was found that active circularly polarized imaging can increase contrast and range compared to linearly polarized imaging. We have completed an initial investigation of the capability for machine learning methods to reduce the effects of light scattering when imaging through fog. Previously acquired experimental long-wave images were used to train an autoencoder denoising architecture. Overfitting was found to be a problem because of lack of variability in the object type in this data set. The lessons learned were used to collect a well labeled dataset with much more variability using the Tabletop Fog Chamber that will be available for future studies. We have developed several new sensing methods using speckle intensity correlations. First, the ability to image moving objects in fog was shown, establishing that our unique speckle imaging method can be implemented in dynamic scattering media. Second, the speckle decorrelation over time was found to be sensitive to fog composition, implying extensions to fog characterization. Third, the ability to distinguish macroscopically identical objects on a far-subwavelength scale was demonstrated, suggesting numerous applications ranging from nanoscale defect detection to security. Fourth, we have shown the capability to simultaneously image and localize hidden objects, allowing the speckle imaging method to be effective without prior object positional information. Finally, an interferometric effect was presented that illustrates a new approach for analyzing speckle intensity correlations that may lead to more effective ways to localize and image moving objects. All of these results represent significant developments that challenge the limits of the application of speckle imaging and open important application spaces. A theory was developed and simulations were performed to assess the potential transverse resolution benefit of relative motion in structured illumination for radar systems. Results for a simplified radar system model indicate that significant resolution benefits are possible using data from scanning a structured beam over the target, with the use of appropriate signal processing.

58 GEOSCIENCES↗

DOE OSTI2024

Efficient Dimension Reduction of Complex Three-dimensional CO2 Saturation using Deep Learning Models

In the domain of deep learning (DL), dimension reduction is crucial for enhancing training efficiency and mitigating overfitting, particularly when managing complex data such as three-dimensional (3D) saturation data. The 3D saturation data in the context of geological carbon storage (GCS) presents unique challenges due to its inherent sparsity and the abrupt transitions at plume boundaries, known as shock fronts. To address the challenges, we proposed a novel DL framework that integrates dimension reduction with advanced 3D reconstruction techniques. Our model leveraged latent variables derived from 2D average saturation data, offering a robust and efficient solution tailored to the intricate dynamics of 3D saturation fields. The proposed framework can extract the critical features of the high-dimensional data while reducing the variable numbers, which is more tractable for DL models and enhances the model robustness and accuracy. Therefore, it provides a novel approach for modeling and analyses in complex geological scenarios, which finds great potential applications in environmental monitoring and energy storage.

Wang, Hongsheng↗

DOE OSTI2024

ML-based Dimension Reduction Strategies

Deep learning (DL)--based surrogate models have achieved success in various applications in carbon capture and storage (CCS). However, the model training on high-dimensional spaces is computationally expensive and impractical for large-scale and complex geological models, because the models usually contain hundreds of thousands to millions of grid cells, each with a set of parameters. Furthermore, the high cost of generating training data with sufficient variation is another limitation of model training on high-dimensional spaces, which may result in overfitting and reduce the model efficiency and prediction performance. We proposed the workflow incorporating dimension reduction methods and deep learning models, which aim to extract the latent variables of input parameters and output state variables, and then build the mapping function at the latent spaces. The proposed workflow can significantly reduce the computational complexity in solving both forward and inverse problems compared to models trained on high-dimensional spaces. Dimensionality reduction models showed great potential in workflows for fast reservoir simulation, history matching, prior model generation, visualization, and more, ultimately enhancing DL model performance in related SMART Work Packages.

Hosseini, Seyyed↗

DOE OSTI2024

Deep Learning-based Parameterization of Complex 3D CO2 Saturation Data in Large-scale Geological Carbon Storage

In deep learning (DL), dimension reduction plays a pivotal role in improving training efficiency and minimizing overfitting, especially when working with complex datasets like three-dimensional (3D) saturation data. In the context of geological carbon storage (GCS), 3D saturation data introduces unique challenges due to its sparse nature and sharp transitions at plume boundaries, known as shock fronts. To tackle these challenges, we developed a novel DL framework that combines dimension reduction with advanced 3D reconstruction techniques. Our approach utilizes latent variables derived from 2D average saturation fields to efficiently capture the essential features of high-dimensional data while reducing the number of variables. This enhances both the robustness and accuracy of DL models, making the framework more practical for real-world applications. By offering a tailored solution for modeling complex 3D saturation dynamics, this framework holds significant potential for environmental monitoring, energy storage, and other geological applications.

Wang, Hongsheng [University of Texas at Austin]↗

DOE OSTI2025

Development of Machine Learning Algorithm for Pebble Bed Modular Reactor Misuse Detection

The objective of this work was to develop a machine learning ensemble that could assist pebble bed reactor verification by evaluating whether a given pebble circulating through a PBR was normal or anomalous using gamma spectroscopy measurements from a notional PBR burnup measurement system. Using a PBR reference design, data sets of synthetic gamma spectra representative of BUMS measurements of normal and anomalous pebbles that may be used to produce special fissile material were generated to train and test an ML anomaly detection ensemble on two reference scenarios – substitution of normal pebbles with target pebbles for production of Pu or 233 U. The ML ensemble correctly identified all anomalous pebbles in the testing data set, and while perfect ensemble performance is normally indicative of overfitting, it was concluded that significantly lower photon intensity of target pebbles produced distinctly less intense photon spectra to where perfect ensemble performance was expected.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗

DOE OSTI2025

An Advanced Microscopic Energy Consumption Model for Automated Vehicle:Development, Calibration, Verification

The automated vehicle (AV) equipped with the Adaptive Cruise Control (ACC) system is expected to reduce the fuel consumption for the intelligent transportation system. This paper presents the Advanced ACC-Micro (AA-Micro) model, a new energy consumption model based on micro trajectory data, calibrated and verified by empirical data. Utilizing a commercial AV equipped with the ACC system as the test platform, experiments were conducted at the Columbus 151 Speedway, capturing data from multiple ACC and Human-Driven (HV) test runs. The calibrated AA-Micro model integrates features from traditional energy consumption models and demonstrates superior goodness of fit, achieving an impressive 90% accuracy in predicting ACC system energy consumption without overfitting. A comprehensive statistical evaluation of the AA-Micro model's applicability and adaptability in predicting energy consumption and vehicle trajectories indicated strong model consistency and reliability for ACC vehicles, evidenced by minimal variance in RMSE values and uniform RSS distributions. Conversely, significant discrepancies were observed when applying the model to HV data, underscoring the necessity for specialized models to accurately predict energy consumption for HV and ACC systems, potentially due to their distinct energy consumption characteristics.

Ma, Ke↗

DOE OSTI2025

Comparing the predictive capabilities of subjective crackle metrics using leave-one-out cross validation

Leave-one-out cross validation was used to show the capability to predict crackling sound quality based on the use of the standard deviation of the time-varying sharpness and the skewness of the derivative of the pressure waveform. The predictive output of this method implies that the more sophisticated fits used to describe this data were not overfit, but instead exhibited substantial predictive value.

Swift, Stephen Hales [Sandia National Laboratories↗

DOE OSTI2026

RHOD Site - NOAA PSL Wind Retrievals WINDoe / Derived Data

This dataset contains daily NetCDF files with horizontal wind profiles retrieved with the WINDoe retrieval (Gebauer and Bell 2024) at Rhode Island (RHOD). WINDoe retrievals datasets are also available at Nantucket Island (NANT, nant.windoe.z01.c1) and Block Island (BLOC, bloc.windoe.z01.c1). WINDoe is an optimal estimation algorithm to retrieve wind profiles combining multiple instruments. The code is available in this github repository (https://github.com/OAR-atmospheric-observations/WINDoe/tree/main) and the retrieval is described by Gebauer and Bell (2024). WINDoe allows combining the individual datasets and outputs into one profile taking into account the information and uncertainties of each dataset. The use of WINDoe minimizes data gaps and maximizes data availability, compared to using wind profiles from only one of the instruments. The regular height grid eases comparisons to numerical weather prediction models. Code modifications have been made that include reading in WFIP3 specific instruments, averaging Doppler lidar radial velocities at various azimuth angles to avoid overfitting, and allowing the user to define a height grid by the user in the vipfile. The instruments used as input to the retrieval are a radar wind profiler (low- and high resolution mode) providing data in and above the boundary layer, a scanning Doppler lidar usually providing data throughout the boundary layer, a profiling lidar providing data from 50 to 200 m at BLOC and NANT, and from 10 to 280 m at Rhode Island, and a surface tower (4 m at NANT and RHOD and 10 m at BLOC). From the scanning lidars, we used radial velocity measurements at 60 deg elevation angle at six different azimuth angles with a resolution of approximately 30 m along the line of sight and the lowest range gate at approximately 70 m. The wind profiles are retrieved with WINDoe up to 3.74 km with 10 m vertical resolution. The profiles are retrieved every 15 min at BLOC and NANT and every 60 min at RHOD.

17 WIND ENERGY↗

DOE OSTI2026

BLOC Site - NOAA PSL Wind Retrievals WINDoe / Derived Data

This dataset contains daily netcdf files with horizontal wind profiles retrieved with the WINDoe retrieval (Gebauer and Bell 2024) at Block Island (BLOC). WINDoe retrievals datasets are also available at Nantucket Island (NANT, nant.windoe.z01.c1) and Rhode Island (RHOD, rhod.windoe.z01.c1). WINDoe is an optimal estimation algorithm to retrieve wind profiles combining multiple instruments. The code is available in this github repository (https://github.com/OAR-atmospheric-observations/WINDoe/tree/main), and the retrieval is described by Gebauer and Bell (2024). WINDoe allows combining the individual datasets and outputs into one profile taking into account the information and uncertainties of each dataset. The use of WINDoe minimizes data gaps and maximizes data availability, compared to using wind profiles from only one of the instruments. The regular height grid eases comparisons to numerical weather prediction models. Code modifications have been made that include reading in WFIP3 specific instruments, averaging Doppler lidar radial velocities at various azimuth angles to avoid overfitting, and allowing the user to define a height grid by the user in the vipfile. The instruments used as input to the retrieval are a radar wind profiler (low- and high resolution mode) providing data in and above the boundary layer, a scanning Doppler lidar usually providing data throughout the boundary layer, a profiling lidar providing data from 50 to 200 m at BLOC and NANT, and from 10 to 280 m at Rhode Island, and a surface tower (4 m at NANT and RHOD and 10 m at BLOC). From the scanning lidars, we used radial velocity measurements at 60 deg elevation angle at six different azimuth angles with a resolution of approximately 30 m along the line of sight and the lowest range gate at approximately 70 m. The wind profiles are retrieved with WINDoe up to 3.74 km with 10 m vertical resolution. The profiles are retrieved every 15 min at BLOC and NANT and every 60 min at RHOD.

17 WIND ENERGY↗

DOE OSTI2026

NANT Site - NOAA PSL Wind Retrievals WINDoe / Derived Data

This dataset contains daily NetCDF files with horizontal wind profiles retrieved with the WINDoe retrieval (Gebauer and Bell 2024) at Nantucket Island (NANT). WINDoe retrievals datasets are also available at Block Island (BLOC, bloc.windoe.z01.c1) and Rhode Island (RHOD, rhod.windoe.z01.c1). WINDoe is an optimal estimation algorithm to retrieve wind profiles combining multiple instruments. The code is available in this github repository (https://github.com/OAR-atmospheric-observations/WINDoe/tree/main), and the retrieval is described by Gebauer and Bell (2024). WINDoe allows combining the individual datasets and outputs into one profile taking into account the information and uncertainties of each dataset. The use of WINDoe minimizes data gaps and maximizes data availability, compared to using wind profiles from only one of the instruments. The regular height grid eases comparisons to numerical weather prediction models. Code modifications have been made that include reading in WFIP3 specific instruments, averaging Doppler lidar radial velocities at various azimuth angles to avoid overfitting, and allowing the user to define a height grid by the user in the vipfile. The instruments used as input to the retrieval are a radar wind profiler (low- and high resolution mode) providing data in and above the boundary layer, a scanning Doppler lidar usually providing data throughout the boundary layer, a profiling lidar providing data from 50 to 200 m at BLOC and NANT, and from 10 to 280 m at Rhode Island, and a surface tower (4 m at NANT and RHOD and 10 m at BLOC). From the scanning lidars, we used radial velocity measurements at 60 deg elevation angle at six different azimuth angles with a resolution of approximately 30 m along the line of sight and the lowest range gate at approximately 70 m. The wind profiles are retrieved with WINDoe up to 3.74 km with 10 m vertical resolution. The profiles are retrieved every 15 min at BLOC and NANT and every 60 min at RHOD.

17 WIND ENERGY↗

DOE OSTI2021

Deep Learning Approaches to Surrogates for Solving the Diffusion Equation for Mechanistic Real-World Simulations

In many mechanistic medical, biological, physical, and engineered spatiotemporal dynamic models the numerical solution of partial differential equations (PDEs), especially for diffusion, fluid flow and mechanical relaxation, can make simulations impractically slow. Biological models of tissues and organs often require the simultaneous calculation of the spatial variation of concentration of dozens of diffusing chemical species. One clinical example where rapid calculation of a diffusing field is of use is the estimation of oxygen gradients in the retina, based on imaging of the retinal vasculature, to guide surgical interventions in diabetic retinopathy. Furthermore, the ability to predict blood perfusion and oxygenation may one day guide clinical interventions in diverse settings, i.e., from stent placement in treating heart disease to BOLD fMRI interpretation in evaluating cognitive function (Xie et al., 2019; Lee et al., 2020). Since the quasi-steady-state solutions required for fast-diffusing chemical species like oxygen are particularly computationally costly, we consider the use of a neural network to provide an approximate solution to the steady-state diffusion equation. Machine learning surrogates, neural networks trained to provide approximate solutions to such complicated numerical problems, can often provide speed-ups of several orders of magnitude compared to direct calculation. Surrogates of PDEs could enable use of larger and more detailed models than are possible with direct calculation and can make including such simulations in real-time or near-real time workflows practical. Creating a surrogate requires running the direct calculation tens of thousands of times to generate training data and then training the neural network, both of which are computationally expensive. Often the practical applications of such models require thousands to millions of replica simulations, for example for parameter identification and uncertainty quantification, each of which gains speed from surrogate use and rapidly recovers the up-front costs of surrogate generation. We use a Convolutional Neural Network to approximate the stationary solution to the diffusion equation in the case of two equal-diameter, circular, constant-value sources located at random positions in a two-dimensional square domain with absorbing boundary conditions. Such a configuration caricatures the chemical concentration field of a fast-diffusing species like oxygen in a tissue with two parallel blood vessels in a cross section perpendicular to the two blood vessels. To improve convergence during training, we apply a training approach that uses roll-back to reject stochastic changes to the network that increase the loss function. The trained neural network approximation is about 1000 times faster than the direct calculation for individual replicas. Because different applications will have different criteria for acceptable approximation accuracy, we discuss a variety of loss functions and accuracy estimators that can help select the best network for a particular application. We briefly discuss some of the issues we encountered with overfitting, mismapping of the field values and the geometrical conditions that lead to large absolute and relative errors in the approximate solution.

60 APPLIED LIFE SCIENCES↗

DOE OSTI2023

Performance Comparison of Object Detection Networks for Shrapnel Identification in Ultrasound Images

Ultrasound imaging is a critical tool for triaging and diagnosing subjects but only if images can be properly interpreted. Unfortunately, in remote or military medicine situations, the expertise to interpret images can be lacking. Machine-learning image interpretation models that are explainable to the end user and deployable in real time with ultrasound equipment have the potential to solve this problem. We have previously shown how a YOLOv3 (You Only Look Once) object detection algorithm can be used for tracking shrapnel, artery, vein, and nerve fiber bundle features in a tissue phantom. However, real-time implementation of an object detection model requires optimizing model inference time. Here, we compare the performance of five different object detection deep-learning models with varying architectures and trainable parameters to determine which model is most suitable for this shrapnel-tracking ultrasound image application. We used a dataset of more than 16,000 ultrasound images from gelatin tissue phantoms containing artery, vein, nerve fiber, and shrapnel features for training and evaluating each model. Every object detection model surpassed 0.85 mean average precision except for the detection transformer model. Overall, the YOLOv7tiny model had the higher mean average precision and quickest inference time, making it the obvious model choice for this ultrasound imaging application. Other object detection models were overfitting the data as was determined by lower testing performance compared with higher training performance. In summary, the YOLOv7tiny object detection model had the best mean average precision and inference time and was selected as optimal for this application. Next steps will implement this object detection algorithm for real-time applications, an important next step in translating AI models for emergency and military medicine.

60 APPLIED LIFE SCIENCES↗