Researchers at the Massachusetts General Hospital (MGH) Department of Radiology have developed an AI system capable of classifying different forms of intracranial hemorrhage using relatively small data sets.

A possible tool for assessing patients with symptoms of potentially life-threatening strokes, the system is capable of producing a quick diagnosis and explaining the reasoning behind it, offering confidence to facilities which may not have access to specially trained neuroradiologists, and addressing the so-called “black box” challenge, in which systems are unable to explain how they arrived at a decision.

“In medicine, it is especially hard to collect high-quality big data. It is critical to have multiple experts label a data set to ensure consistency of data. This process is very expensive and time-consuming. Some critics suggest that machine learning algorithms cannot be used in clinical practice, because the algorithms do not provide justification for their decisions. We realized that it is imperative to overcome these two challenges to facilitate the use in health care of machine learning, which has an immense potential to improve the quality of and access to care," authors Sehyo Yune, the director of researcher translation at MGH Radiology, and Hyunkwang Lee, a graduate student at the Harvard School of Engineering and Applied Sciences, said in a statement.

The FDA requires that all decision support systems provide data for users to review the reasoning behind their findings.

Trained on 904 head CT scans, the MGH solution classifies images into one of five hemorrhage subtypes, based on the location of the brain, or as no hemorrhage, reviewing and saving images from the training data set that most accurately represent the traditional features found in each classification. It then uses an atlas of distinguishing features to show a group of images similar to those of the CT scan under evaluation to explain its diagnosis.

The team designed the system to mimic the way in which radiologists analyze images, incorporating adjusting factors such as contrast and brightness for revealing subtle differences that are not immediately apparent, and the ability to scroll through adjacent CT scans to determine whether or not a finding on a single image reflects a real problem or is a meaningless artifact.

Following its completion, the researchers tested the system on a retrospective set of 100 scans with and 100 scans without intracranial hemorrhages that were taken before the system was developed, and on a prospective set that was taken after its creation of 79 scans with and 117 without hemorrhages.

The system proved to be just as accurate as radiologists in identifying and classifying intracranial hemorrhages from the retrospective set, and even better than non-expert human readers in its assessment of the prospective one.

In addition, the system can be deployed directly onto the scanner, allowing it to alert care teams to the presence of a hemorrhage for appropriate further testing to take place before the patient is even off the scanner.

“The next step will be to deploy the system into clinical areas and further validate its performance with many more cases,” said author Shahein Tajmir, a radiology resident at MGH Radiology. “We are currently building a platform to allow for the widespread application of such tools throughout the department. Once we have this running in the clinical setting, we can evaluate its impact on turnaround time, clinical accuracy and the time to diagnosis."

Each of the 904 heat CT scans used to train the system consisted of around 40 individual images that were labeled by five MGH neuroradiologists.

Partial support was provided through a grant by the National Institutes of Health.

The findings were published online in the journal, Nature Biomedical Engineering.