Mass General builds algorithm to detect intracranial hemorrhages

An artificial intelligence algorithm has been developed by Massachusetts General Hospital to rapidly diagnose and classify brain hemorrhages from unenhanced head computed-tomography scans.

The deep learning system not only detects acute intracranial hemorrhage—and classifies five subtypes of the condition—but generates predictions from the data that are explainable, according to researchers, who contend that the tool could potentially help hospital emergency departments to evaluate patients experiencing acute stroke symptoms.

Brain Images-CROP.jpg

“Owing to improvements in image recognition via deep learning, machine-learning algorithms could eventually be applied to automated medical diagnoses that can guide clinical decision-making,” according to study results published in the journal Nature Biomedical Engineering. “Our approach to algorithm development can facilitate the development of deep-learning systems for a variety of clinical applications and accelerate their adoption into clinical practice.”

Also See: Machine learning helps Geisinger cut time to diagnose intracranial hemorrhages

“Rapid recognition of intracranial hemorrhage, leading to prompt appropriate treatment of patients with acute stroke symptoms, can prevent or mitigate major disability or death,” says co-author Michael Lev, MD, director of emergency radiology and emergency neuroradiology at MGH.

“Many facilities do not have access to specially trained neuroradiologists—especially at night or over weekends—which can require non-expert providers to determine whether or not a hemorrhage is the cause of a patient’s symptoms,” adds Lev. “The availability of a reliable, ‘virtual second opinion’—trained by neuroradiologists—could make those providers more efficient and confident and help ensure that patients get the right treatment.”

A team from MGH’s Department of Radiology trained the AI algorithm using 904 head CT scans and then tested it on two separate datasets.

“The system achieved a performance similar to that of expert radiologists in two independent test datasets containing 200 cases (sensitivity of 98 percent and specificity of 95 percent) and 196 cases (sensitivity of 92 percent and specificity of 95 percent),” states the study. “The system includes an attention map and a prediction basis retrieved from training data to enhance explainability, and an iterative process that mimics the workflow of radiologists.”

For reprint and licensing requests for this article, click here.