Associate Professor University of Delaware, United States
Introduction: Magnetic resonance elastography (MRE) is a phase-contrast MRI technique which images wave patterns in tissue produced by induced mechanical vibrations in order to non-invasively quantify tissue mechanical properties, such as shear stiffness. MRE can detect changes in mechanical properties related to various neurological disorders [1], and whole brain maps of in vivo mechanical properties are useful in building computational models of brain biomechanics for simulating individual responses to head impacts in traumatic brain injury [2]. However, brain MRE remains primarily a research technique, with the need for additional hardware and specialized imaging sequences making it not yet widely available. As such, there are only a limited number of available brain images for producing comprehensive models of brain biomechanical response to understand traumatic brain injury across subjects [3]. Data augmentation techniques hold promise for generating artificial brain mimics that reflect the subtle features of real brain stiffness maps, which can be used to provide biomechanical data necessary to build advanced computational models in the absence of subject-specific MRE data. The goal of this work was to develop a machine learning approach to generate artificial brain stiffness mimics from training on a series of available real images in order to augment currently existing datasets.
Materials and
Methods: The model developed is a variant of a Wasserstein Generative Adversarial Network with Lipschitz Penalty (WGAN-LP) [4] adapted to allow for the creation of three-dimensional images and was trained on a dataset of 83 brain stiffness maps registered to the MNI152 template that were selected from the MRE134 atlas of healthy subjects aged 18 to 35 [5]. Training of the model makes use of two adversarial networks: a generator and a critic (Figure 1). During training, the generator was given inputs of random noise, which were progressively upscaled and convolved to generate novel images of identical dimension to that of the real brain images. The critic was given sets of images where half of the sets would be drawn from the real dataset and the remaining sets of images were created by the generator. The critic produces a “realness” score for the input generated images through successive convolution layers to progressively downsample and process the images. The realness scores from the critic were then used in the respective loss functions of the two networks to iteratively refine the networks and further a competition where the generator produces successively improved images, while the critic is more able to distinguish between the real and generated images. This training process continued until the entirety of the training dataset was iterated across 3000 times to train the generator network. Subsequently, a batch of 83 brain mimics were generated so that they could be compared to the training dataset of the same size.
Results, Conclusions, and Discussions: Figure 2 shows a series of generated brain stiffness mimics, which strongly resembled those of real stiffness maps, with common features such as the ventricles resolving similarly. Subsequently, the example set of generated images and the entirety of the training dataset were quantitatively compared to each other to determine how well the GAN learned to mimic actual data. All images were eroded with a cuboid kernel of width 2 to limit edge effects, and each image set was averaged together to produce the mean image for each group, where again there was strong agreement between input and generated images (Figure 3). Additionally, every image was compared to the true subject mean using the structural similarity index measure (SSIM), a measure of image agreement. Compared with the true mean image, both input (0.776 0.016) and generated images (0.776 0.031) had good agreement; the range of SSIMs was not significantly different between sets (p = 0.891, 2 < 0.01). These metrics are desired, where the distribution of generated images is consistent with the population distribution, but also with a variety of maps capturing subject variability. Overall, we demonstrate that a WGAN-LP can be used to generate stiffness mimics which are similar to real stiffness maps, capture the general topology of the brain, and demonstrate variability like those found in real subjects, despite the limited size and scope of the input dataset. Future work will involve increasing the overall size of the training dataset to better capture variation in brain properties, and increase the age range of the input training data. Additionally, work will focus on introducing the ability to differentiate between different groups of subjects, allowing for the creation of brain mimics similar to subjects of a specific demographic instead of the overall population.
Acknowledgements (Optional): This work was supported in part by NIH grant U01-NS112120.
[1] M. C. Murphy, J. Huston, and R. L. Ehman, “MR elastography of the brain and its application in neurological diseases,” NeuroImage, vol. 187. Academic Press Inc., pp. 176–183, Feb. 15, 2019. doi: 10.1016/j.neuroimage.2017.10.008. [2] P. V. Bayly et al., “MR Imaging of Human Brain Mechanics In Vivo: New Measurements to Facilitate the Development of Computational Models of Brain Injury,” Annals of Biomedical Engineering, vol. 49, no. 10. Springer, pp. 2677–2692, Oct. 01, 2021. doi: 10.1007/s10439-021-02820-0. [3] A. Alshareef et al., “Integrating material properties from magnetic resonance elastography into subject-specific computational models for the human brain,” Brain Multiphys, vol. 2, Jan. 2021, doi: 10.1016/j.brain.2021.100038. [4] H. Petzka, A. Fischer, and D. Lukovnicov, “On the regularization of Wasserstein GANs,” Sep. 2018, [Online]. Available: http://arxiv.org/abs/1709.08894 [5] L. V. Hiscox et al., “Standard-space atlas of the viscoelastic properties of the human brain,” Hum Brain Mapp, vol. 41, no. 18, pp. 5282–5300, Dec. 2020, doi: 10.1002/hbm.25192.
