A team of researchers from the Department of Radiology at Massachusetts General Hospital (MGH) developed a system using artificial intelligence to quickly diagnose and classify brain haemorrhages and to provide the basis for their decisions from relatively small sets of imaging data. Such a system could become an indispensable tool for hospital emergency departments that evaluate patients with symptoms of a potentially life-threatening stroke, allowing rapid application of the correct treatment. The team report was published online in Biomedical engineering of nature.
Although increasing computing power and the availability of large data sets improve machine learning – the process by which computers analyze data, identify patterns, and learn essentially how to perform a task without the direct involvement of a human programmer – important obstacles may prevent systems from being integrated into clinical decision making. This includes the need for large, well-annotated datasets – pre-developed image analysis systems capable of duplicating a physician's performance have been trained with more than 100,000 images – and the "black box" problem, explain how they came to a decision. The US Food and Drug Administration requires that any decision support system provide data that allows users to analyze the reasons behind their findings.
"It's somewhat paradoxical to use the words" small data "or" explainable "to describe a study that used deep learning," says Hyunkwang Lee, a graduate student at the Faculty of Engineering and Applied Science at Harvard, one of the two main authors. of the study. "However, in medicine, it is especially difficult to collect large, high-quality data. It is critical that a number of experts rank a set of data to ensure data consistency, which is very costly and time-consuming."
Dr. Sehyo Yune of MGH Radiology adds, "Some critics suggest that machine learning algorithms can not be used in clinical practice because algorithms do not provide justification for their decisions. It is imperative to overcome these two challenges to facilitate the health use of machine learning, which has immense potential for improving quality and access to care. "
To train their system, the MGH team began with 904 head CT scans, each consisting of about 40 individual images, which were labeled by a team of five MGH neuroradiologists if they described one of five subtypes of bleeding, with based on the location within the brain, or without bleeding. To improve the accuracy of this deep learning system, the team – led by senior author Synho Do, PhD, director of the Laboratory of Radiology of Imaging and Computation MGH and an assistant professor of Radiology at Harvard Medical School – built in steps that mimic the path radiologists analyze images. These include adjustment factors such as contrast and brightness to reveal subtle differences not immediately apparent and scroll through adjacent computed tomography slices to determine whether something appearing in a single image reflects a real problem or is a meaningless artifact.
Once the model system was created, the researchers tested it on two separate sets of CT scans – a retrospective set made before system development, consisting of 100 scans with and 100 scans without intracranial hemorrhage, and a prospective set of 79 scans with and without haemorrhage, taken after the creation of the model. In their analysis of the retrospective set, the model system was as accurate in the detection and classification of intracranial hemorrhages as the radiologists who reviewed the CT scans. In his analysis of the prospective set, it proved to be even better than unspecialized human readers.
To solve the "black box" problem, the team reviewed the system and saved the images from the training data set that more clearly represented the classic characteristics of each of the five subtypes of bleeding. Using this atlas of distinctive features, the system is able to display a group of images similar to those of the computed tomography being analyzed in order to explain the basis of its decisions.
"The rapid recognition of intracranial hemorrhage, leading to an appropriate treatment of patients with acute stroke symptoms, can prevent or lessen major disability or death," says co-author Michael Lev, MD, MGH Radiology. "Many facilities do not have access to specially trained neuroradiologists – especially at night or on weekends – which may require non-specialist providers to determine whether or not a bleeding is the cause of a patient's symptoms. "opinion – trained by neuroradiologists – could make these providers more efficient and confident and help ensure that patients receive the correct treatment. "
MGH Radiology adds: "In addition to providing the necessary second opinion, this system can also be deployed directly to scanners, alerting the attending staff to bleeding and triggering appropriate additional tests before The next step will be to deploy the system in clinical areas and further validate its performance with many other cases.We are currently building a platform to enable the broad application of these tools across the department.We have this working in the clinical setting, we can assess its impact response time, clinical accuracy and time to diagnosis. "