The Massachusetts General Radiology Department inspectors team developed a system using artificial intelligence to quickly diagnose and classify cerebral palsy and provide a basis for solutions from relatively small image sets. Such a system can be an indispensable tool for hospital emergency services that evaluate patients with potentially life-threatening symptoms. The team report was published online Biomedical Engineering in Nature.
Always increasing the computational power and great data availability of computers – computers analyze data, define patterns, and teach how to perform a task without being directly involved in a single programmer – important obstacles can be avoided. Integration of systems into clinical decision making. Among them there is a need for a great and well-illustrated database – more than 100,000 images of previously developed image analysis systems allowing a repeat of a doctor's work and the problem of "black box", and how they can explain the systems. The US Food and Drug Administration requires any decision-support system to provide users with the information that enables them to review the causes behind the findings.
Hyunkwang Lee, a postgraduate student at the Harvard Engineering and Applied Science School, is one of the two leading writers, Hyunkwang Lee, a little paradoxical to learning about "small data" or "perception." research. "However, it is particularly difficult to collect high-quality information in medicine, and it is important that most experts have a database to ensure the accuracy of the data, which is very expensive and often takes too long."
Sehyo Yune, MD author of MRI radiology, says: "Some critics think that machine learning algorithms can not be used in clinical practice, as some critics do not justify algorithms, and we understand that it is important to eliminate these two problems, and there are problems to improve the health of machine knowledge, has a great potential to achieve. "
To prepare the system, the MGH group started with 904 CT scanning, consisting of 40 individual images, each of which describes one of the five bleeding subgroups of the five MGH neuroscientists. There is no brain or bleeding. In order to increase the accuracy of this in-depth learning system, the head of the MGH Radiology Laboratory, PhD, Medical Sciences and Radiology Associate at Harvard Medical School, under the guidance of senior researcher Sinho Doun. These include factors that correlate with contrast and brightness to reveal incomprehensible and subtle differences with adjacent CT scanning samples to determine whether something in the image is a real problem or a meaningless artifact.
After the model system was created, the investigators tested two CT components – a retrospective set before the system preparation, 100 scanning and 100 scanning capabilities without intracranial hemorrhage, and 79 scanning and 117 bleeding followed by model acquisition. In the retrospective kit analysis, the modeling system was accurate in detecting and classifying intracranial hemorrhages as well as radiologists with scanners. In the predicted set of analysis, non-expert people proved to be better than human readers.
To solve the "black box" problem, the team reviewed the system and retrieved images from training data that reflected the classic features of each of the five bleeding subtypes. Using this satin with distinctive features, the system can illustrate a group of analogues similar to the CT scan to explain the basis of the decisions.
"Rapid recognition of intracranial hemorrhage can prevent or alleviate severe disability or death when administered to patients with acute stroke symptoms," says co-author Michael Lev, MD, MGH Radiology. "Many devices may not be able to utilize specialized neuroradiology, especially night or weekend, to require non-specialist providers to determine if a bleeding patient is the cause of the symptoms.The ideas taught by neuro-radiologists can make these providers more efficient and confident can provide proper treatment of patients. "
Co-rapporteur Shahein Tacir, MD, MGH Radiology, added: "This is a very useful virtual second thought, and this system can also be placed directly on the scanners, warns a bleeding entity, and then triggers relevant tests. The patient is away from the scanner, clinical sites, and, in many cases, its functionality, we are currently building a platform that enables the distribution of such tools in the department, and we can experience this disease and we can evaluate its time, clinical accuracy and diagnosis time. "