University of Wisconsin School of Medicine and Public Health
Machine learning can make an impact on patients recovering from intracerebral hemorrhage
Intracerebral hemorrhage (ICH), or bleeding in the brain, affects over 100,000 people in this country each year. Bleeding strokes, though less common than strokes from blockages, tend to result more often in death or long-term extended care. To reduce the degree of long-term care needed after ICH, care teams need to be able to respond quickly and precisely in removing the hemorrhaged blood.
A team at the University of Wisconsin-Madison has developed a machine learning network using MRI images to aid neurosurgeons in this process.
Similar to a stroke in its long-term effects, people with an ICH can experience a profound reduction in quality of life and require long-term extended care. However, a national trial showed that long-term patient independence was improved when doctors were able to remove the hemorrhaged blood from the brain so that 15 milliliters or less remained.
This careful balancing act requires eliminating exactly enough blood in the correct locations while simultaneously avoiding a potential reinjury to the brain. To accomplish this, neurosurgeons need to be able to quickly visualize the relationship between the interventional device, the remaining blood volume and its edges. A key finding in the national trial also showed that current tools were not meeting this need, and automating the visualization and quantification of the injured brain could prove beneficial to surgeons at various decision points.
Recently published in the journal Magnetic Resonance Imaging, lead author Thomas Lilieholm, MS, PhD student in the Department of Medical Physics, developed and trained a machine learning network to automatically classify whole-brain MRI scans into color-coded overlays identifying blood clot and edema over regions of normal brain. These overlays output precise, milliliter measurements of clot volumes and denote the spread, shape, and location of hemorrhaged blood.
The network was built using retrospective, locally-acquired MRI patient scans and were then compared to three independently generated sets of overlaid images created by team members Matt Henningsen (Department of Electrical & Computer Engineering), radiology fellow Dr. Matt Larson, and Lilieholm. While CT scans are currently more commonly used for ICH, this network incorporated MR-imaging because of its superior ability to evaluate damaged soft tissue compared to CT.
Results showed that the machine learning approach was able to identify and segment volumes of remaining blood and edema in high agreement with human observers. With three sets of human-produced segmentations to compare against, the model is robust in its ability to match varied manual standards.
Going forward, the team sees several areas for future development. The accuracy of the machine learning will continue to be improved as the team collects more patient MRI data. This also demonstrates the value of MRI in treating ICH and the potential for more specific applications in pathology segmentation. Most importantly, this work is available for consideration as neurosurgeons work to design the next upcoming national clinical trial in ICH removal.
“This cutting-edge technology is precisely what we require to expertly navigate minimally invasive procedures for the most perilous form of stroke – hemorrhagic or bleeding strokes,” says UW Health neurosurgeon Dr. Azam Ahmed, a past surgical participant in national ICH evacuation trials. “This automated system enhances every facet of the stroke team’s operations, enabling us to pinpoint the ideal path for hemorrhage removal while closely monitoring its impact on brain swelling.”
The team on this project includes a collaboration between several departments at UW-Madison.
Dr. Azam Ahmed, associate professor of neurological surgery and radiology. Medical director of the University of Wisconsin Comprehensive Brain Tumor Program.
Walter Block, PhD, professor biomedical engineering, medical physics & radiology
Matthew Henningsen, MS, department of electrical & computer engineering
Matthew Larson, MD, PhD, fellow, department of radiology
Thomas Lilieholm, PhD student, department of medical physics
Alan McMillan, PhD, professor of radiology & medical physics