Massachusetts General’s AI can spot brain hemorrhages as accurately as humans

Intracranial hemorrhages — brain bleeds — are the second-most common subtype of stroke, according to the National Center for Biotechnology Information. They interrupt blood flow around or inside of the brain, depriving it of oxygen, which is why timely treatment is critical. After three or four minutes, brain cells begin to die.
Fortunately, artificial intelligence (AI) promises to drive progress on the diagnostic front. In a paper (“An explainable deep-learning algorithm for the detection of acute intracranial hemorrhages from small datasets”) published in the journal Nature Biomedical Engineering last month, researchers at Massachusetts General Hospital in Boston describe a deep learning algorithm that can detect acute intracerebral hemorrhages, or ICHs, with a high degree of accuracy.
Their work comes about a month after researchers at the University of California, Berkeley demonstrated an AI system that can predict Alzheimer’s disease from brain scans up to six years in advance.
“It is somewhat paradoxical to use the words ‘small data’ or ‘explainable’ to describe a study that used deep learning,” Hyunkwang Lee, a graduate student at the Harvard School of Engineering and Applied Sciences and one of the two lead authors of the study, told MedicalXpress. “However, in medicine, it is especially hard to collect high-quality big data. It is critical to have multiple experts label a dataset to ensure consistency of data. This process is very expensive and time-consuming.”
The researchers trained their AI system using a dataset of 904 computerized tomography (CT) scans — scans that combine a series of X-ray images taken from different angles — of patients’ heads, each comprising about 40 individual images. A five-person team of neuroradiologists labeled each, identifying those that depicted one of five hemorrhage subtypes.
As the model ingested each scan, it performed transformations to improve the quality of analyses. For example, it adjusted the contrast and brightness of images, and examined adjacent scan slices to sort artifacts from signs of hemorrhage.
In tests performed with two independent datasets — a retrospective set of 100 scans with and 100 without ICH, and a prospective set of 79 scans with and 117 without hemorrhage — the AI system performed quite well. On the retrospective set, it performed on par with expert radiologists. And in its evaluation of the second dataset — the prospective set — it achieved better accuracy than non-expert humans.
The model’s other novelty is its explainable design. In an effort to make transparent the rationale behind each of its decisions, the team had the AI system review and highlight scans from the training dataset that clearly represented the features of each of the five hemorrhage subtypes. Predictions the algorithm makes are accompanied by groups of images similar those of the scan it analyzed.
Dr. Michael Lev, one of the paper’s coauthors, said that AI systems like it could one day provide trained clinicians with a second opinion, or step in during days at facilities and care providers when a trained neuroradiologist isn’t available.
“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,” he said.
Shortly, the team plans to deploy the system in clinical areas and further validate its performance with more cases. Coauthor Dr. Shahein Tajmir said that work at Massachusetts General has already begun on a platform to allow for the “widespread application of such tools.”
Source: VentureBeat