A team of researchers from the Massachusetts General Medical Department (MGH) Department of Radiology has developed a system using artificial information to diagnose and distribute brain hemorrhages quickly and to provide a basis for the decisions of relatively small image data sets. Such a system could become an indispensable tool for hospital emergency departments evaluating patients with stroke symptoms that could pose a life threat, allowing the correct treatment to be applied quickly. The report of the online team was published in Aberystwyth Nature of Biomedical Engineering.
Although ever increasing accounting power and the availability of large datasets have improved learning machines – the process in which computers analyze data, identify patterns and essentially teach themselves how to perform a task without direct involvement in programming human – important barriers can prevent such systems to be integrated into clinical decision-making. These include the need for large data sets and good annotations – previously developed imaging analysis systems were able to duplicate the performance of a trained doctor with more than 100,000 images – a & # 39; r "black box" problem, inability of systems to explain how they reach a decision. The US Food and Drug Administration requires any decision-making system to provide data that allows users to review the reasons behind their findings.
"It's a bit paradoxical to use the words" small data "or explained to describe a study that used deep learning," said Hyunkwang Lee, a graduate at School Engineering and Applied Sciences Harvard, one of the two leading authors of the study. "However, in medicine, it is particularly difficult to collect high quality data. It is essential that multiple experts label a dataset to ensure data consistency .This process is very expensive and take a lot of time. "
Co-conductor, Sehyo Yune, MD, from MGH Raddoleg, added: "Some critics suggest that machine learning algorithms can not be used in clinical practice, as algorithms do not provide justification for their decisions. We realized it It is essential to overcome both challenges to facilitate the use of plant learning healthcare, which have huge potential to improve quality and access to care. "
To train their system, the MGH team started with 904 head CT scans, each containing about 40 individual images, labeled by a team of five MGH neurrobiologists about whether they depicted one of five hemorrhage isipipes, based on the location of within the brain, or without hemorrhage. In order to improve the accuracy of this deep learning system, the team – led by senior author Synho Do, PhD, director of the MGH Movement Medicine Radiography Laboratory and Medical Computing and Assistant Radiology professor at Harvard Medical School – have build in steps that imitate the way in which radiologists analyze images. These include modifying factors such as contrast and brightness to reveal subtle differences that are not immediately visible and scroll through adjacent CT scissors to determine whether something that appears on a single image reflects a problem that ok or not to be a meaningless artefact.
Once the model system was created, the researchers were tested on two sets of separate CT scans – taken retrospectively before the system was developed, including 100 scans with 100 without intracranial hemorrhage, and A set of 79 scans with 117 without hemorrhage, which was taken after the model was created. In its retrospective set analysis, the model system was so accurate in detecting and distributing intracranial hemorrhages as the radiologists had reviewed the scans. In its analysis of the proposed set, they were even better than non-expert man readers.
To solve the "black box" problem, the team had the system review and rescue the images of the training data set that clearly demonstrated the classic features of each of the five hemorrhage isipipes. Using this atlas of different features, the system can show a group of similar images to those whose CT scan is analyzed to explain the basis of its decisions.
"Rapid recognition of intracranial hemorrhage, which results in appropriate treatment appropriately for patients with acute stroke symptoms, can prevent or mitigate a disability or a great death," said co-author Michael Lev, MD , MGH Radiology. "Many facilities do not have access to specially trained neuroradiologists – especially during the evenings or over weekends – who may not be required for non-specialist providers to determine whether hemorrhage is a cause Patient symptoms The virtual availability of reliable opinions – have been trained by neuroradiologists – enables those providers to be more efficient and confident and help to ensure patients get the right treatment. "
Co-author Shahein Tajmir, MD, MGH Radiology adds, "As well as providing that second opinion, this system can be used directly on scanners, warning & # 39; The care team attends hemorrhage presence and triggers appropriate further tests before the patient is even off the scanner. The next step will be to use the system in clinical areas and further validate its performance with many more cases. We are currently building a platform to allow the application of such tools across the department, we have this running in the clinical setting, we can evaluate its impact on turning time, clinical accuracy and time to get diagnosed. "