The Big Data Institute research team built a software test that automatically diagnoses diseases based on input data of X-ray images, with 90% accuracy.
The competition to diagnose pneumothorax was held online with more than 1,000 participating teams. Photo: BNE
The software is built on open data from actual radiographs of both people with and without disease, then for machine learning and recognition. The results show that if based on the evaluation criteria of the medical industry, both the sensitivity and specificity of the algorithm reach over 90%, even over 95% for some diseases. If based on the criteria of processing speed, to read a standard X-ray image will need 5-10 minutes but the reader only takes 2-5 seconds depending on the hardware configuration.
Professor Vu Ha Van, Director of the Big Data Research Institute (Big Data Institute – Vingroup) said: “In order to be sure of our ability to conduct research of international standards, the Analysis team The Institute’s Medical Imaging has submitted its software to participate in 2 prestigious competitions, held on an open platform, in the competition to diagnose pneumothorax (Pneumothorax Segmentation) organized by the American Association for Medical Imaging Informatics. (SIIM), the Institute’s research team is in the top 5 positions out of a total of more than 1,000 competing teams from all over the world, including many strong teams from prestigious organizations of Vietnam. Russia, USA, China, Israel…
And in the competition to diagnose 12 common lung diseases through chest radiographs (CheXpert) organized by Stanford University – USA, the group is also ranked 5th out of 40 teams. In particular, the algorithm’s diagnostic results were better or comparable when compared with the panel of radiologists organized by Stanford to evaluate the competition, on a specific set of radiographs.”
According to Dr. Nguyen Quy Ha, Team Leader of Medical Image Analysis (Big Data Institute), despite initial positive results, the algorithm is being built based on foreign open data sources. If it is possible to collect medical image data of Vietnamese people, the accuracy and applicability in practice in Vietnam will be really high.
Mr. Ha said that for each disease, the algorithm needs to take pictures from 100,000 to 200,000 patients. “The bigger the input data, the better. Just like the more diverse a doctor is exposed to, the more experience and accurate the diagnosis,” Dr. Ha said, adding that this is a difficulty due to the fact that Standard medical data collection cannot be done in day one, day two.
To prepare for input data, on July 23, Big Data Institute signed a Memorandum of Understanding with the Information Technology Department – Ministry of Health. The two sides agreed to coordinate in developing regulations on the exchange of medical examination and treatment data of people between medical examination and treatment facilities and organizations and individuals for research and development. products applying artificial intelligence in Health.
“The data after being normalized and fully annotated will not carry any personal information and will be shared widely with the community. When the software is complete, hospitals can be used for free. Doctors will receive maximum support, minimizing time to read images.In lower-level hospitals, where there are not many specialists, the application of AI will significantly increase the opportunity for patients to have access to Modern diagnostic methods help doctors have more useful information to decide on therapy,” said Professor Vu Ha Van.
When the product is completed, one possible option is to deploy a remote diagnostic system that applies cloud computing technology. Users will be granted an account and just send an image to have the reading results sent back soon after. This approach has the advantage that it is cheap and hospitals do not need to invest in their own infrastructure.
To put the software to use, in addition to collecting Vietnamese people’s data, legal issues and licensing procedures will be the next steps being taken. “We will aim at the necessary diagnoses for some dangerous and common diseases in Vietnam such as cancers or cardiovascular, neurological and diabetes diseases. The aim is that the software must achieve the target. international standards and become useful to domestic hospitals,” said Prof. Vu Ha Van.
According to experts, the AI algorithm helps to increase the accuracy of image diagnosis by detecting many hidden features in the image, which are not similar to what the naked eye sees. AI can help localize questionable lesions on images early so doctors can focus on them and improve diagnosis. AI makes consistent and tireless diagnoses, is capable of synthesizing a huge amount of information in images, and relates previous case diagnosis histories, which is difficult to follow with conventional methods. old.