Real-time dynamic simulation for highly accurate spatiotemporal brain deformation from impact

A high-throughput framework for predicting three-dimensional structural-mechanical relationships of human cranial bones using a deep learning-based method

Cranial bone injuries significantly impact human health, potentially leading to death or permanent disability, and mechanical responses are crucial predictors of cranial damage. Predicting mechanical responses through medical imaging is an efficient method that streamlines the process by eliminating the need for intermediate steps, such as diagnostic testing and biomedical analysis. Although previous studies have successfully predicted 1D sequences or 2D sectional mechanical attributes from medical imaging, these approaches are limited in their ability to capture the anisotropic characteristics of cranial bone. As a complex osseous material, cranial bone has significant directional dependencies in its microstructural features, which directly influence its macroscopic mechanical responses under loading conditions. In this study, we aim to introduce a deep learning-based high-throughput framework to correlate the three-dimensional mechanical responses and three-dimensional cranial microstructures derived from medical images. First, micro-CT scans of 40 human cranial samples, spanning an average age of 82.5 years, were performed to capture microstructural information. Next, 2000 representative volume element (RVE) units were randomly and automatically extracted from these scans to characterize the cranial microstructures. Following this, 2000 stress and 2000 strain fields were derived from finite element simulations based on these RVE units, and subsequently validated through quasi-static compression experiments. An optimized U-Net-based deep learning network was employed to link the macro-property stress-strain response with the three-dimensional cranial microstructures. The proposed framework demonstrates robust performance in predicting the spatial mechanical behavior based on microstructural inputs, showing high and consistent similarity between the predictions and ground truth. Overall, the high-throughput nature of this framework facilitates the handling of large-scale data, enabling comprehensive and efficient analysis that is crucial for predicting mechanical responses. By elucidating structure-property relationships, this approach can enhance the accuracy of injury diagnosis and aid in the development of tailored treatment plans, effectively bridging the gap between structural morphology and mechanical functionality.

Learning soft tissue deformation from incremental simulations

BACKGROUND

Surgical planning for orthognathic procedures demands swift and accurate biomechanical modeling of facial soft tissues. Efficient simulations are vital in the clinical pipeline, as surgeons may iterate through multiple plans. Biomechanical simulations typically use the finite element method (FEM). Prior works divide FEM simulations into increments to enhance convergence and accuracy. However, this practice elongates simulation time, thereby impeding clinical integration. To accelerate simulations, deep learning (DL) models have been explored. Yet, previous efforts either perform simulations in a single step or neglect the temporal aspects in incremental simulations.

PURPOSE

This study investigates the use of spatiotemporal incremental modeling for biomechanics simulations of facial soft tissue.

METHODS

We implement the method using a graph neural network. Our method synergizes spatial features with temporal aggregation using DL networks trained on incremental FEM simulations from 17 subjects that underwent orthognathic surgery.

RESULTS

Our proposed spatiotemporal incremental method achieved a mean accuracy of 0.37 mm with a mean computation time of 1.52 s. In comparison, a spatial-only incremental method yielded a mean accuracy of 0.44 mm and a mean computation time of 1.60 s, while a spatial-only single-step method yielded a mean accuracy of 0.41 mm and a mean computation time of 0.05 s.

CONCLUSIONS

Statistical analysis demonstrated that the spatiotemporal incremental method reduced mean errors compared to the spatial-only incremental method, emphasizing the importance of incorporating temporal information in incremental simulations. Overall, we successfully implemented spatiotemporal incremental learning tailored to simulate soft tissue deformation while substantially reducing simulation time compared to FEM.

Peaks and Distributions of White Matter Tract-related Strains in Bicycle Helmeted Impacts: Implication for Helmet Ranking and Optimization

Traumatic brain injury (TBI) in cyclists is a growing public health problem, with helmets being the major protection gear. Finite element head models have been increasingly used to engineer safer helmets often by mitigating brain strain peaks. However, how different helmets alter the spatial distribution of brain strain remains largely unknown. Besides, existing research primarily used maximum principal strain (MPS) as the injury parameter, while white matter fiber tract-related strains, increasingly recognized as effective predictors for TBI, have rarely been used for helmet evaluation. To address these research gaps, we used an anatomically detailed head model with embedded fiber tracts to simulate fifty-one helmeted impacts, encompassing seventeen bicycle helmets under three impact locations. We assessed the helmet performance based on four tract-related strains characterizing the normal and shear strain oriented along and perpendicular to the fiber tract, as well as the prevalently used MPS. Our results showed that both the helmet model and impact location affected the strain peaks. Interestingly, we noted that different helmets did not alter strain distribution, except for one helmet under one specific impact location. Moreover, our analyses revealed that helmet ranking outcome based on strain peaks was affected by the choice of injury metrics (Kendall's Tau coefficient: 0.58-0.93). Significant correlations were noted between tract-related strains and angular motion-based injury metrics. This study provided new insights into computational brain biomechanics and highlighted the helmet ranking outcome was dependent on the choice of injury metrics. Our results also hinted that the performance of helmets could be augmented by mitigating the strain peak and optimizing the strain distribution with accounting the selective vulnerability of brain subregions and more research was needed to develop region-specific injury criteria.

Differences between two maximal principal strain rate calculation schemes in traumatic brain analysis with in-vivo and in-silico datasets

Brain deformation caused by a head impact leads to traumatic brain injury (TBI). The maximum principal strain (MPS) was used to measure the extent of brain deformation and predict injury, and the recent evidence has indicated that incorporating the maximum principal strain rate (MPSR) and the product of MPS and MPSR, denoted as MPS × SR, enhances the accuracy of TBI prediction. However, ambiguities have arisen about the calculation of MPSR. Two schemes have been utilized: one is to use the time derivative of MPS (MPSR1), and another is to use the first eigenvalue of the strain rate tensor (MPSR2). Both MPSR1 and MPSR2 have been applied in previous studies to predict TBI. To quantify the discrepancies between these two methodologies, we compared them across eight in-vivo and one in-silico head impact datasets and found that 95MPSR1 was slightly larger than 95MPSR2 and 95MPS × SR1 was 4.85 % larger than 95MPS × SR2 in average. Across every element in all head impacts, the average MPSR1 was 12.73 % smaller than MPSR2, and MPS × SR1 was 11.95 % smaller than MPS × SR2. Furthermore, logistic regression models were trained to predict TBI using MPSR (or MPS × SR), and no significant difference was observed in the predictability. The consequence of misuse of MPSR and MPS × SR thresholds (i.e. compare threshold of 95MPSR1 with value from 95MPSR2 to determine if the impact is injurious) was investigated, and the resulting false rates were found to be around 1 %. The evidence suggested that these two methodologies were not significantly different in detecting TBI.

On the application of hybrid deep 3D convolutional neural network algorithms for predicting the micromechanics of brain white matter

BACKGROUND

Material characterization of brain white matter (BWM) is difficult due to the anisotropy inherent to the three-dimensional microstructure and the various interactions between heterogeneous brain-tissue (axon, myelin, and glia). Developing full scale finite element models that accurately represent the relationship between the micro and macroscale BWM is however extremely challenging and computationally expensive. The anisotropic properties of the microstructure of BWM computed by building unit cells under frequency domain viscoelasticity comprises of 36 individual constants each, for the loss and storage moduli. Furthermore, the architecture of each unit cell is arbitrary in an infinite dataset.

METHODS

In this study, we extend our previous work on developing representative volume elements (RVE) of the microstructure of the BWM in the frequency domain to develop 3D deep learning algorithms that can predict the anisotropic composite properties. The deep 3D convolutional neural network (CNN) algorithms utilizes a voxelization method to obtain geometry information from 3D RVEs. The architecture information encoded in the voxelized location is employed as input data while cross-referencing the RVEs' material properties (output data). We further develop methods by incorporating parallel pathways, Residual Neural Networks and inception modulus that improve the efficiency of deep learning algorithms.

RESULTS

This paper presents different CNN algorithms in predicting the anisotropic composite properties of BWM. A quantitative analysis of the individual algorithms is presented with the view of identifying optimal strategies to interpret the combined measurements of brain MRE and DTI.

SIGNIFICANCE

The proposed Multiscale 3D ResNet (M3DR) algorithm demonstrates high learning ability and performance over baseline CNN algorithms in predicting BWM tissue properties. The hybrid M3DR framework also overcomes the significant limitations encountered in modeling brain tissue using finite elements alone including those such as high computational cost, mesh and simulation failure. The proposed framework also provides an efficient and streamlined platform for implementing complex boundary conditions, modeling intrinsic material properties and imparting interfacial architecture information.

Efficient Generation of Pretraining Samples for Developing a Deep Learning Brain Injury Model via Transfer Learning

The large amount of training samples required to develop a deep learning brain injury model demands enormous computational resources. Here, we study how a transformer neural network (TNN) of high accuracy can be used to efficiently generate pretraining samples for a convolutional neural network (CNN) brain injury model to reduce computational cost. The samples use synthetic impacts emulating real-world events or augmented impacts generated from limited measured impacts. First, we verify that the TNN remains highly accurate for the two impact types (N = 100 each;R 2 of 0.948-0.967 with root mean squared error, RMSE, ~ 0.01, for voxelized peak strains). The TNN-estimated samples (1000-5000 for each data type) are then used to pretrain a CNN, which is further finetuned using directly simulated training samples (250-5000). An independent measured impact dataset considered of complete capture of impact event is used to assess estimation accuracy (N = 191). We find that pretraining can significantly improve CNN accuracy via transfer learning compared to a baseline CNN without pretraining. It is most effective when the finetuning dataset is relatively small (e.g., 2000-4000 pretraining synthetic or augmented samples improves success rate from 0.72 to 0.81 with 500 finetuning samples). When finetuning samples reach 3000 or more, no obvious improvement occurs from pretraining. These results support using the TNN to rapidly generate pretraining samples to facilitate a more efficient training strategy for future deep learning brain models, by limiting the number of costly direct simulations from an alternative baseline model. This study could contribute to a wider adoption of deep learning brain injury models for large-scale predictive modeling and ultimately, enhancing safety protocols and protective equipment.

Potential of Soft-Shelled Rugby Headgear to Lower Regional Brain Strain Metrics During Standard Drop Tests

BACKGROUND

The growing concern for player safety in rugby has led to an increased focus on head impacts. Previous laboratory studies have shown that rugby headgear significantly reduces peak linear and rotational accelerations compared to no headgear. However, these metrics may have limited relevance in assessing the effectiveness of headgear in preventing strain-based brain injuries like concussions. This study used an instantaneous deep-learning brain injury model to quantify regional brain strain mitigation of rugby headgear during drop tests. Tests were conducted on flat and angled impact surfaces across different heights, using a Hybrid III headform and neck.

RESULTS

Headgear presence generally reduced the peak rotational velocities, with some headgear outperforming others. However, the effect on peak regional brain strains was less consistent. Of the 5 headgear tested, only the newer models that use open cell foams at densities above 45 kg/m3 consistently reduced the peak strain in the cerebrum, corpus callosum, and brainstem. The 3 conventional headgear that use closed cell foams at or below 45 kg/m3 showed no consistent reduction in the peak strain in the cerebrum, corpus callosum, and brainstem.

CONCLUSIONS

The presence of rugby headgear may be able to reduce the severity of head impact exposure during rugby. However, to understand how these findings relate to brain strain mitigation in the field, further investigation into the relationship between the impact conditions in this study and those encountered during actual gameplay is necessary.

Effect of Impact Kinematic Filters on Brain Strain Responses in Contact Sports

OBJECTIVE

Impact kinematics are widely employed to investigate mechanisms of traumatic brain injury (TBI). However, they are susceptible to noise and artefacts; thus, require data filtering. Few studies have focused on how data filtering affects brain strain most relevant to TBI. Here, we report that impact-induced brain strains are much less sensitive to data filtering than kinematics based on three filtering methods: CFC180, lowpass 200 Hz, and a new method called Head Exposure to Acceleration Database in Sport (HEADSport).

METHODS

Using mouthguard-measured head impacts in elite rugby (N = 5694), average Euclidean distances between the three filtered angular velocity profiles and their unfiltered counterparts are used to identify three groups of impacts with large variations: 90-95th, 95-99th, and >99th percentile. From each group, 20 impacts are randomly selected for simulation using the anisotropic Worcester Head Injury Model (WHIM) V1.0.

RESULTS AND CONCLUSION

HEADSport and CFC180 are the most and least effective, respectively, in suppressing "unphysical artefacts" shown as sharp spikes with a rather short impulse duration (e.g., <3 ms) in angular velocity. However, maximum principal strain (MPS), especially that in the corpus callosum, is much less sensitive to data filtering compared to kinematic peaks (e.g., reduction of 3% vs. 47% and 90% for peak angular velocity and acceleration with HEADSport for impacts of >99th percentile).

SIGNIFICANCE

These findings confirm that the brain acts as a low-pass filter, itself, to suppress high frequency noise in impact kinematics. Therefore, brain strain can serve as a common metric for TBI biomechanical studies to maximize relevance to the injury, as it is not sensitive to kinematic filters.

Brain Deformation Estimation With Transfer Learning for Head Impact Datasets Across Impact Types

OBJECTIVE

The machine-learning head model (MLHM) to accelerate the calculation of brain strain and strain rate, which are the predictors for traumatic brain injury (TBI), but the model accuracy was found to decrease sharply when the training/test datasets were from different head impacts types (i.e., car crash, college football), which limits the applicability of MLHMs to different types of head impacts and sports. Particularly, small sizes of target dataset for specific impact types with tens of impacts may not be enough to train an accurate impact-type-specific MLHM.

METHODS

To overcome this, we propose data fusion and transfer learning to develop a series of MLHMs to predict the maximum principal strain (MPS) and maximum principal strain rate (MPSR).

RESULTS

The strategies were tested on American football (338), mixed martial arts (457), reconstructed car crash (48) and reconstructed American football (36) and we found that the MLHMs developed with transfer learning are significantly more accurate in estimating MPS and MPSR than other models, with a mean absolute error (MAE) smaller than 0.03 in predicting MPS and smaller than [Formula: see text] in predicting MPSR on all target impact datasets. High performance in concussion detection was observed based on the MPS and MPSR estimated by the transfer-learning-based models.

CONCLUSION

The MLHMs can be applied to various head impact types for rapidly and accurately calculating brain strain and strain rate.

SIGNIFICANCE

This study enables developing MLHMs for the head impact type with limited availability of data, and will accelerate the applications of MLHMs.

A computational pipeline towards large-scale and multiscale modeling of traumatic axonal injury

Contemporary biomechanical modeling of traumatic brain injury (TBI) focuses on either the global brain as an organ or a representative tiny section of a single axon. In addition, while it is common for a global brain model to employ real-world impacts as input, axonal injury models have largely been limited to inputs of either tension or compression with assumed peak strain and strain rate. These major gaps between global and microscale modeling preclude a systematic and mechanistic investigation of how tissue strain from impact leads to downstream axonal damage throughout the white matter. In this study, a unique subject-specific multimodality dataset from a male ice-hockey player sustaining a diagnosed concussion is used to establish an efficient and scalable computational pipeline. It is then employed to derive voxelized brain deformation, maximum principal strains and white matter fiber strains, and finally, to produce diverse fiber strain profiles of various shapes in temporal history necessary for the development and application of a deep learning axonal injury model in the future. The pipeline employs a structured, voxelized representation of brain deformation with adjustable spatial resolution independent of model mesh resolution. The method can be easily extended to other head impacts or individuals. The framework established in this work is critical for enabling large-scale (i.e., across the entire white matter region, head impacts, and individuals) and multiscale (i.e., from organ to cell length scales) modeling for the investigation of traumatic axonal injury (TAI) triggering mechanisms. Ultimately, these efforts could enhance the assessment of concussion risks and design of protective headgear. Therefore, this work contributes to improved strategies for concussion detection, mitigation, and prevention.

Adaptive Machine Learning Head Model Across Different Head Impact Types Using Unsupervised Domain Adaptation and Generative Adversarial Networks

Machine learning head models (MLHMs) are developed to estimate brain deformation from sensor-based kinematics for early detection of traumatic brain injury (TBI). However, the overfitting to simulated impacts and the decreasing accuracy caused by distributional shift of different head impact datasets hinders the broad clinical applications of current MLHMs. We propose a new MLHM configuration that integrates unsupervised domain adaptation with a deep neural network to predict whole-brain maximum principal strain (MPS) and MPS rate (MPSR). With 12,780 simulated head impacts, we performed unsupervised domain adaptation on target head impacts from 302 college football (CF) impacts and 457 mixed martial arts (MMA) impacts using domain regularized component analysis (DRCA) and cycle-GAN-based methods. The new model improved the MPS/MPSR estimation accuracy, with the DRCA method outperforming other domain adaptation methods in prediction accuracy: MPS mean absolute error (MAE): 0.017 (CF) and 0.020 (MMA); MPSR MAE: 4.09 s-1 (CF) and 6.61 s-1(MMA). On another two hold-out test sets with 195 college football impacts and 260 boxing impacts, the DRCA model outperformed the baseline model without domain adaptation in MPS and MPSR estimation MAE. The DRCA domain adaptation approach reduces the error of MPS/MPSR estimation to be well below previously reported TBI thresholds, enabling accurate brain deformation estimation to detect TBI in future clinical applications.

A Morphologically Individualized Deep Learning Brain Injury Model

The brain injury modeling community has recommended improving model subject specificity and simulation efficiency. Here, we extend an instantaneous (< 1 sec) convolutional neural network (CNN) brain model based on the anisotropic Worcester Head Injury Model (WHIM) V1.0 to account for strain differences due to individual morphological variations. Linear scaling factors relative to the generic WHIM along the three anatomical axes are used as additional CNN inputs. To generate training samples, the WHIM is randomly scaled to pair with augmented head impacts randomly generated from real-world data for simulation. An estimation of voxelized peak maximum principal strain of the whole-brain is said to be successful when the linear regression slope and Pearson's correlation coefficient relative to directly simulated do not deviate from 1.0 (when identical) by more than 0.1. Despite a modest training dataset (N = 1363 vs. ∼5.7 k previously), the individualized CNN achieves a success rate of 86.2% in cross-validation for scaled model responses, and 92.1% for independent generic model testing for impacts considered as complete capture of kinematic events. Using 11 scaled subject-specific models (with scaling factors determined from pre-established regression models based on head dimensions and sex and age information, and notably, without neuroimages), the morphologically individualized CNN remains accurate for impacts that also yield successful estimations for the generic WHIM. The individualized CNN instantly estimates subject-specific and spatially detailed peak strains of the entire brain and thus, supersedes others that report a scalar peak strain value incapable of informing the location of occurrence. This tool could be especially useful for youths and females due to their anticipated greater morphological differences relative to the generic model, even without the need for individual neuroimages. It has potential for a wide range of applications for injury mitigation purposes and the design of head protective gears. The voxelized strains also allow for convenient data sharing and promote collaboration among research groups.

A sensor-enabled cloud-based computing platform for computational brain biomechanics

BACKGROUND AND OBJECTIVES

Driven by the risk of repetitive head trauma, sensors have been integrated into mouthguards to measure head impacts in contact sports and military activities. These wearable devices, referred to as "instrumented" or "smart" mouthguards are being actively developed by various research groups and organizations. These instrumented mouthguards provide an opportunity to further study and understand the brain biomechanics due to impact. In this study, we present a brain modeling service that can use information from these sensors to predict brain injury metrics in an automated fashion.

METHODS

We have built a brain modeling platform using several of Amazon's Web Services (AWS) to enable cloud computing and scalability. We use a custom-built cloud-based finite element modeling code to compute the physics-based nonlinear response of the intracranial brain tissue and provide a frontend web application and an application programming interface for groups working on head impact sensor technology to include simulated injury predictions into their research pipeline.

RESULTS

The platform results have been validated against experimental data available in literature for brain-skull relative displacements, brain strains and intracranial pressure. The parallel processing capability of the platform has also been tested and verified. We also studied the accuracy of the custom head surfaces generated by Avatar 3D.

CONCLUSION

We present a validated cloud-based computational brain modeling platform that uses sensor data as input for numerical brain models and outputs a quantitative description of brain tissue strains and injury metrics. The platform is expected to generate transparent, reproducible, and traceable brain computing results.

Approximating subject-specific brain injury models via scaling based on head-brain morphological relationships

Most human head/brain models represent a generic adult male head/brain. They may suffer in accuracy when investigating traumatic brain injury (TBI) on a subject-specific basis. Subject-specific models can be developed from neuroimages; however, neuroimages are not typically available in practice. In this study, we establish simple and elegant regression models between brain outer surface morphology and head dimensions measured from neuroimages along with age and sex information (N = 191; 141 males and 50 females with age ranging 14-25 years). The regression models are then used to approximate subject-specific brain models by scaling a generic counterpart, without using neuroimages. Model geometrical accuracy is assessed using adjusted [Formula: see text] and absolute percentage error (e.g., 0.720 and 3.09 ± 2.38%, respectively, for brain volume when incorporating tragion-to-top). For a subset of 11 subjects (from smallest to largest in brain volume), impact-induced brain strains are compared with those from "morphed models" derived from neuroimage-based mesh warping. We find that regional peak strains from the scaled subject-specific models are comparable to those of the morphed counterparts but could be considerably different from those of the generic model (e.g., linear regression slope of 1.01-1.03 for gray and white matter regions versus 1.16-1.19, or up to ~ 20% overestimation for the smallest brain studied). These results highlight the importance of incorporating brain morphological variations in impact simulation and demonstrate the feasibility of approximating subject-specific brain models without neuroimages using age, sex, and easily measurable head dimensions. The scaled models may improve subject specificity for future TBI investigations.

Use of Brain Biomechanical Models for Monitoring Impact Exposure in Contact Sports

Head acceleration measurement sensors are now widely deployed in the field to monitor head kinematic exposure in contact sports. The wealth of impact kinematics data provides valuable, yet challenging, opportunities to study the biomechanical basis of mild traumatic brain injury (mTBI) and subconcussive kinematic exposure. Head impact kinematics are translated into brain mechanical responses through physics-based computational simulations using validated brain models to study the mechanisms of injury. First, this article reviews representative legacy and contemporary brain biomechanical models primarily used for blunt impact simulation. Then, it summarizes perspectives regarding the development and validation of these models, and discusses how simulation results can be interpreted to facilitate injury risk assessment and head acceleration exposure monitoring in the context of contact sports. Recommendations and consensus statements are presented on the use of validated brain models in conjunction with kinematic sensor data to understand the biomechanics of mTBI and subconcussion. Mainly, there is general consensus that validated brain models have strong potential to improve injury prediction and interpretation of subconcussive kinematic exposure over global head kinematics alone. Nevertheless, a major roadblock to this capability is the lack of sufficient data encompassing different sports, sex, age and other factors. The authors recommend further integration of sensor data and simulations with modern data science techniques to generate large datasets of exposures and predicted brain responses along with associated clinical findings. These efforts are anticipated to help better understand the biomechanical basis of mTBI and improve the effectiveness in monitoring kinematic exposure in contact sports for risk and injury mitigation purposes.

Piecewise Multivariate Linearity Between Kinematic Features and Cumulative Strain Damage Measure (CSDM) Across Different Types of Head Impacts

In a previous study, we found that the relationship between brain strain and kinematic features cannot be described by a generalized linear model across different types of head impacts. In this study, we investigate if such a linear relationship exists when partitioning head impacts using a data-driven approach. We applied the K-means clustering method to partition 3161 impacts from various sources including simulation, college football, mixed martial arts, and car crashes. We found piecewise multivariate linearity between the cumulative strain damage (CSDM; assessed at the threshold of 0.15) and head kinematic features. Compared with the linear regression models without partition and the partition according to the types of head impacts, K-means-based data-driven partition showed significantly higher CSDM regression accuracy, which suggested the presence of piecewise multivariate linearity across types of head impacts. Additionally, we compared the piecewise linearity with the partitions based on individual features used in clustering. We found that the partition with maximum angular acceleration magnitude at 4706 rad/s² led to the highest piecewise linearity. This study may contribute to an improved method for the rapid prediction of CSDM in the future.

American Football Helmet Effectiveness Against a Strain-Based Concussion Mechanism

Brain strain is increasingly being used in helmet design and safety performance evaluation as it is generally considered as the primary mechanism of concussion. In this study, we investigate whether different helmet designs can meaningfully alter brain strains using two commonly used metrics, peak maximum principal strain (MPS) of the whole brain and cumulative strain damage measure (CSDM). A convolutional neural network (CNN) that instantly produces detailed brain strains is first tested for accuracy for helmeted head impacts. Based on N = 144 impacts in 12 impact conditions from three random and representative helmet models, we conclude that the CNN is sufficiently accurate for helmet testing applications, for elementwise MPS (success rate of 98.6%), whole-brain peak MPS and CSDM (coefficient of determination of 0.977 and 0.980, with root mean squared error of 0.015 and 0.029, respectively). We then apply the technique to 23 football helmet models (N = 1104 impacts) to reproduce elementwise MPS. Assuming a concussion would occur when peak MPS or CSDM exceeds a threshold, we sweep their thresholds across the value ranges to evaluate the number of predicted hypothetical concussions that different helmets sustain across the impact conditions. Relative to the 12 impact conditions tested, we find that the "best" and "worst" helmets differ by an average of 22.5% in terms of predicted concussions, ranging from 0 to 42% (the latter achieved at the threshold value of 0.28 for peak MPS and 0.4 for CSDM, respectively). Such a large variation among helmets in strain-based concussion predictions demonstrate that helmet designs can still be optimized in a clinically meaningful way. The robustness and accuracy of the CNN tool also suggest its potential for routine use for helmet design and safety performance evaluation in the future. The CNN is freely available online at https://github.com/Jilab-biomechanics/CNN-brain-strains .