Machine Learning for Prediction of Fetal Neurodevelopmental Abnormalities
Machine Learning for Prediction of Fetal Neurodevelopmental Abnormalities
Isabelle Deng Thomas Jefferson High School for Science and Technology
This article was originally included in the 2020 print publication of the Teknos Science Journal.
July 12, 2019, 7:00 am. Surrounded by a group of radiologists, technicians, and neonatologists, I sat in a large magnetic resonance imaging (MRI) room furnished with eight active screens and communication consoles. Toscanned through the hemodynamic readings and dozens of annotated 3D fetal brain images we were processing through the active console. This was one of the first fetuses whose prenatal images were compiled; it was also part of the Integrated Research Information System (IRIS), our lab’s neuroimaging web platform for the real-time monitoring of fetal neurodevelopment and diagnostics of congenital disorders. I headed back to my cubicle in the adjacent room and peered at my desktop screen, monitoring the progress of the machine learning network training on the over 1500 neuroimaging and time-series data. In just one hour, we were ready for the first phase of clinical testing.
About 15 million infants around the world are born prematurely, and more than 40,000 are afflicted with complex congenital heart disease (CHD) [1]. Although recent advances in the management of CHD have reduced mortality, brain injury remains a major impediment to high quality of life among survivors. A growing body of literature has suggested that much of this neurological morbidity may be prenatal in origin. A study in 2018 established that CHD fetuses show impaired brain development, predominantly in the third trimester of pregnancy which is associated with impaired brain perfusion [1]. Even though qualitative MRI studies have revealed structural patterns associated with disrupted neurodevelopment, the altered functional brain connectivity has not been extensively studied. Thus, the goal of my internship at the Center for the Developing Brain at Children’s National Hospital was to analyze functional MRI (fMRI) images and create a computational tool that can predict and identify signatures of complex CHD-associated abnormalities as well as assess fetal brain maturity. We hope to utilize and streamline novel machine learning (ML) and computer vision algorithms for eventual noninvasive monitoring of fetal neurodevelopment of preterm and CHD fetuses at multiple time scales, bridging yet a new barrier in fetal medicine research.
Machine learning (ML) is based on the idea that systems can learn from data, identify patterns, and make decisions with minimal human intervention. Deep neural networks and generative adversarial networks, which imitate the workings of the human brain in processing data, have spurred discussion in the computer vision community because they have proven to be efficient in medical image processing tasks like image reconstruction, detection, and data augmentation. Within medical imaging, ML methods are highly successful in gleaning descriptive features of brain structure from non-invasive measurements and in recovering features that differentiate patients and controls.
One of the first studies that attempted to integrate machine learning and fetal imaging was conducted at the Wayne State University Neuroimaging Lab.T Researchers implemented a 3D Convolutional Neural Network (CNN) in conjunction with CNN interpretation techniques for the analysis of fetal brain fMRI to predict the fetus’s gestational age [2]. Their model was able to predict gestational age with an 84% accuracy rate, which surpassed that of traditional supervised techniques such as the Random Forest and Decision Tree ML models [3]. By visualizing the network during training, Li et al. (2019) were able to identify the brain regions, including the thalamus and cerebellum, whose signals play a defining role in differentiating between younger and older fetuses. This study was an a stepping stone for my own research, which involves a comparison between the performance of CNN and ensemble learners to effectively model fetal growth trajectories and identify biomarkers seen in fetuses with CHD.
Currently there are major impediments in the utilization of machine learning for processing neuroimages, one of which is the lack of annotated data. Many algorithms, including deep neural networks, are highly data-dependent and require copious amounts of training data to effectively distinguish patterns in similar images. In medical imaging tasks, image annotations made by expert radiologists are time-consuming, especially for precise annotations such as the segmentation of the fetal brain or lesions in 3D image volumes. To address the lack of high-quality functional brain images, researchers at the Beckman Institute of the University of Illinois conducted a study that demonstrated the efficacy of using generative models as a tool for synthesizing artificial, high-quality, diverse, and task-dependent functional brain images, improving the performance of predictive models for the classification of liver lesions [5]. Sanmi Koyejo (personal communication, Jan. 18, 2020) also said that using ensemble models with generative models and visualizing the network layers would enable a close comparison of the functional images of the case and control subjects to identify subtle brain network differences. Thus, utilizing generative models to introduce rich variability into the dataset will be an important part of my CHD predictive project.
In addition, even though advances in quantitative imaging tools have improved noninvasive monitoring, the presence of fetal brain artifacts, noise due to motion in fMRI images, and the low quality of the image dataset have been major impediments for data interpretation by ML models. In a pioneering study, researchers from the University of Michigan and Yale School of Medicine aimed to implement a semi-automated resting-state fMRI processing pipeline that streamlined processing steps such as motion denoising, volume realignment, normalization, and fetal brain segmentation from maternal tissue by utilizing convolutional neural networks [4]. They attained a 92% segmentation accuracy and were able to run the full pipeline on 360 fMRI volumes in 13.2 minutes on the GPU [4]. Thus, incorporating and optimizing an automated image processing pipeline is an essential component of my research to yield high quality, motion-free fMRI images to use during the “learning” process of a deep learning model for CHD prediction.
Predicting the presence of complex CHD developmental biomarkers and designing a computational platform as a neuroimaging data visualization hub will afford clinicians and patients a new window to understand neurodevelopment at the earliest stages of life. The early identification and diagnosis of brain-related problems may help us develop treatments to prevent them as early as possible and ensure the best possible outcomes for babies with congenital disorders.
References
[1] De Asis-Cruz, J., Donofrio, M. T., Vezina, G., & Limperopoulos, C. (2018). Aberrant brain functional connectivity in newborns with congenital heart disease before cardiac surgery. Neuroimage: Clinical, 17, 1-10. https://doi.org/10.1016/j.nicl.2017.09.020
[2] Li, X., Hect, J., Thomason, M., & Zhu, D. (2019). Interpreting age effects of human fetal brain from spontaneous fMRI using deep 3D convolutional neural networks. Neurocomputing, 3, 15-25. https://doi.org/arXiv:1906.03691
[3] Luo, Y., Li, Z., Guo, H., Cao, H., Song, C., Guo, X., & Zhang, Y. (2017). Predicting congenital heart defects: A comparison of three data mining methods. PLOS ONE, 12(5), 1-20. https://doi.org/10.1371/journal.pone.0177811
[4] Rutherford, S., Sturmfels, P., Angstadt, M., Hect, J., Wiens, J., Huevel, M., . . . Sripada, C. (2019). Observing the origins of human brain development: Automated processing of fetal fMRI. Neurocomputing, 4, 10-27. https://doi.org/10.1101/525386
[5] Zhuang, P., Schwing, A. G., & Koyejo, O. (2019). FMRI data augmentation via synthesis. 2019 IEEE 16th International Symposium on Biomedical Imaging, 1, 1-13. https://doi.org/10.1109/ISBI.2019.8759585
