AI-Driven Radiomics and Genomics for Predicting Brain Health
Artificial intelligence (AI) and machine learning are increasingly making important contributions to biomedical research. Neurology and radiology are where these tools are poised to have a major impact, particularly in developing applications to aid in interpreting MRI scans and other brain images.
Researchers at Mass General Brigham are at the forefront of this work.
"By using machine learning and other algorithms to study brain images, it's possible to gain much more information than you could ever get with the human eye," says Natalia Rost, MD, MPH, chief of the Stroke Division in the Department of Neurology at Massachusetts General Hospital. "When we digitize these images and turn them into 0s and 1s, we can identify characteristics on grayscale, texture, and granularity that we otherwise couldn't even fathom."
Radiomics-derived Brain Age Suggests Clinical Outcomes
Dr. Rost's lab is using these tools to develop ways to predict how patients will fare after having a stroke. One recent study published by her group in Neurology focused not on traditional chronological age but on what she calls radiomics-derived brain age.
"Across an individual's lifetime, exposures to high blood pressure, diabetes, and hypercholesterolemia, as well as lifestyle factors like smoking, can have a big impact on the health of the brain and recovery after an acute stroke," she says. "Patients and their doctors may not be aware that they have these silent changes, but the preexisting burden of disease in the brain could be a marker of how well patients will recover."
Through large international collaborations, Dr. Rost and her colleagues have created a repository of brain scans housed at Mass General Brigham. Working with colleagues in the Computer Science and Artificial Intelligence Laboratory (CSAIL) at Massachusetts Institute of Technology, she and her team have utilized this repository to develop a machine learning algorithm that quantifies brain age and shows potential for predicting outcomes after stroke.
The research demonstrated that T2 MRI detectable changes reflect cardiovascular risk factor accumulation and are robustly linked to worse outcomes after stroke.
"The algorithms validate some of the things we already observe in the clinic," Dr. Rost says. Eventually, she adds, it may be possible to implement these tools in clinical practice and use them to guide treatment and early rehabilitation decisions for patients with acute ischemic stroke.
Amassing Data to Develop Deep Learning Models
Another focus of Dr. Rost's lab is using deep learning to predict stroke severity. She and her team published a paper in Brain Communications comparing the capacities of classically used linear algorithms and novel deep learning algorithms for these predictions. The models leveraged convolutional neural networks to determine how much data is needed for deep learning models to outperform traditional methods.
"Deep learning is considered to be a 'black box,' and scientists have traditionally been skeptical and apprehensive about using this approach because of the lack of transparency about how these technologies actually work," she says. "We decided to take a deep dive into understanding what it takes for a deep learning model to match or even overcome the performance of regular statistical algorithms."
The study found the deep learning model starts to outperform other algorithms at nine-fold the amount of data. She explains that while using these deep learning models is much further from clinical practice than the machine learning models, the ultimate goal is the same: developing tools to predict patient outcomes after stroke and eventually guide clinical decisions.
Tools Aid in Understanding Interface of Vascular and Nervous Systems
These tools also have implications for studying brain age and brain health more broadly, including in understanding brain changes associated with conditions like dementia and Parkinson's disease.
"Neurodegeneration has multiple forms," Dr. Rost says. "Injury that happens during or immediately after a stroke can precipitate changes in the brain that ultimately look like Alzheimer's disease. These tools can help us understand the interface between the vascular and nervous systems and its role in developing brain pathology—or, conversely, in brain resilience."
She and her team hope to create brain health biomarkers one day that will be applicable across a wide range of neurologic conditions.
Developing Prediction Models to Molecularly Diagnose Brain Tumors
AI and machine learning could advance molecular diagnosis of glioblastoma. Predictive models are starting to make it possible to determine a tumor's molecular subtype using MRI.
"Machine learning provides additional value in interpreting these MRIs and identifying features that are not easily discernible by the human eye," says Raymond Y. Huang, MD, PhD, division chief of Neuroradiology at Brigham and Women's Hospital.
Dr. Huang and his team have built models that can determine whether a glioblastoma tumor is IDH mutant or wildtype and also diagnose its 1p/19q co-deletion status.
"These models allow us to utilize what's already available—the MRI images—and extract more information from them to enhance our diagnosis," he explains.
In an era of increased reliance on molecular diagnosis of tumors, understanding the genomic status of glioblastoma tumors is important not only for prognosis but also for guiding surgical planning and making other treatment decisions. Increasingly, drugs are available to target tumors that carry these genetic alterations. But for brain tumors, obtaining tumor tissue surgically for analysis can be challenging, making it a priority to find other noninvasive ways to diagnose the molecular mutations of these tumors.
"It helps to know a tumor's molecular status in advance to determine the extent of the surgery that will be performed," Dr. Huang says. "In the future, it may even be possible to use some of these new targeted therapies in the neoadjuvant setting to shrink tumors before surgery and make them easier to remove."
Collaboration Leads to More Training Data
Collaboration with other centers around the world is an important part of developing these models, according to Dr. Huang.
"These tumors are rare, and even at a center like ours that treats a relatively large number of patients, it can be hard to collect enough data points to train these models," he says. "When multiple centers work together and share data, it further enhances our ability to get accurate models for predicting tumor genetics."
Dr. Huang adds that working with other institutions also ensures these models are generalizable across centers.
"Differences in how imaging and data are acquired can cause machine learning to fail," he says. "When we include data collected across different sites, the models are better able to work with these differences."
Deep Learning Helps to Identify Predictive Features
Among the MRI features included in the training data are tumor size, shape, volume, and surface area. The model also looks at increases or decreases at the pixel level and the variation in pixel intensity across the tumor.
Deep learning can identify predictive features directly from the MRI images. In the team's experiments, the three-dimensional heterogeneity of these tumors is modeled using three representative orthogonal slices—on the axial, coronal, and sagittal planes.
The training data is further augmented by introducing random rotations, translations, sharing, zooming, and flipping of the images. The MRI features are correlated with data about tumor histology and molecular status that are collected later.
Dr. Huang notes there are many possible future directions for this research, including seeking regulatory approval for using these prediction models in the clinical setting, expanding the models to genomic characteristics beyond IDH and 1p/19q, and increasing the number of sites participating in the research. He also expects these tools will be useful for assessing glioblastoma response to therapies, including IDH inhibitors.
"Across Mass General Brigham, we have a lot of talented people working together to develop these tools," Dr. Huang says. "It's not just radiology but also individuals from many specialties who care for patients with these aggressive brain tumors. We are all working together to develop these studies. And beyond oncology, there are many opportunities for using AI to diagnose and treat neurological diseases."