What if a pre-operative MRI could indicate whether a glioma was likely to have a genetic mutation that significantly alters prognosis?

Preliminary success in using machine learning to analyze MRI data suggests the promise of artificial intelligence (AI) to create a “virtual biopsy” that would increase pre-operative precision in diagnosis and prognosis of brain tumors, including glioblastoma multiforme (GBM). Omar Arnaout, MD, a neurosurgeon in the Brigham’s Department of Neurosurgery, is leading a research team that has demonstrated that AI can be used to identify isocitrate dehydrogenase (IDH) mutation based on MRI data. IDH mutation, which confers significantly longer survival, was chosen for testing because IDH status may influence treatment plans, perioperative patient counseling and adjuvant management.

“Historically, we’ve needed tissue. Now, we can predict, very reliably just from an MRI, whether the patient has this mutation,” said Arnaout, co-director of the Computational Neuroscience Outcomes Center (CNOC) at Harvard, which focuses on neurosurgical applications of artificial intelligence. “Theoretically, that can guide your conversation with the patient, guide your surgical aggressiveness and guide the kind of treatment.”

Arnaout’s team collaborated with other hospitals to test whether the IDH status of gliomas could be predicted accurately from MRI by applying a residual convolutional neural network to preoperative radiographic data. To do this, the researchers combined preoperative imaging from 496 patients at three hospitals (including the Brigham) and used two patient cohorts to train a machine to differentiate IDH-mutated tumors (associated with longer overall survival) from tumors without the mutation (which have a poor prognosis). Independent performance testing on a third cohort predicted IDH status with 79 to 87 percent accuracy, the authors reported in Clinical Cancer Research.

Creating Virtual Biopsy Software

Building on that foundation, the research team is conducting prospective validation of pre-operative mutation detection in gliomas. This moves them closer to their goal of creating accessible “virtual biopsy” software — with broad implications.

“IDH is only one of many clinically relevant mutations in GBM,” Arnaout said. “We’re taking structural MRI data and using it to predict not only biomarker status in a brain tumor but also clinical outcome and use that as a decision aid.”

Ongoing work combines multiple rich data sets to contribute to machine learning:

  • Images, including perioperative CT and MRI and from post-surgical monitoring
  • “Unstructured data” from electronic medical records, analyzed through natural language processing
  • Pathology, dating back several decades at the Brigham, now being digitized and used as input for machine learning algorithms
  • Physiological data from patients in the Brigham’s Neurosurgical Intensive Care Unit that is being collected and archived for research

Another CNOC initiative, led by CNOC Co-Director Timothy R. Smith, MD, PhD, MPH, involves collecting data outside the context of the health care environment to help predict readmissions and to identify early any tumor recurrence or decline in cognitive function. This digital phenotyping includes tracking smartphone data of patients to assess, from their daily usage, whether they are experiencing subtle cognitive decline.

“AI can be a good tool for balancing treatment effectiveness versus patient harm for each individual case. This can help us focus the difficult conversations around risk and benefit, to optimize quality of life,” said Arnaout.

A version of this story originally appeared in Brigham Vital Lines.