Artificial intelligence in diagnosis of cancer
Main article: Diagnosis of cancer
The artificial intelligence is actively applied in researches of development of techniques of diagnosis of cancer.
As neuronets train to describe x-rays
One of the main tasks of AI is the help to physicians in prevention of medical errors and also removal of inspections on essentially new qualitative level at the expense of the accuracy of data analysis and the description.
Training of a neuronet in radiology happens as follows: before giving start to test use of the program, more than 200,000 x-rays were analyzed and work is continued. A system makes calculation of an error, and further specialists make setup of network. Experienced radiologists with a long standing participate in this action. Directly each x-ray not only is marked, but also 3 experts, independently of each other analyze. If results of researches match at all radiologists, then they are used already afterwards by a neuronet. Further the trained neuronet is connected to the radiological systems. Having received the picture of the patient, AI, on the basis of earlier obtained and modified data on researches, does the conclusion. The doctor, in turn, if necessary adjusts the conclusion and supplements.
Advantages of use of AI in radiology
Advantages of use of a neuronet directed by radiological diagnoses which are noted by the Russian specialists for August, 2019:
- AI allows to describe the x-ray in 3 seconds - the specialist describes the picture up to 20 minutes. In a sheaf "AI doctor" the description of a research makes less than 3 seconds. According to present standards carrying out and the description of one research can take up to 90 minutes.
- Low working costs of a system
- AI allows to be the second opinion for the radiologist. "The second opinion" allows to learn the opinion of the independent expert, to obtain more information on a disease and the plan of treatment. The reliability of diagnostics in this case increases by 48%.
- Also AI allows to be the second opinion for the doctor-clinical physician
- AI helps to separate a flow of patients and prioritization, having removed a considerable part of this loading from the doctor
- Not less important point is the possibility of quality control by means of technology and audit.
- System accuracy in the description of the picture together with the doctor is 95-98%. The neuronet selects specific area on which pathology was found that allows the doctor to draw a conclusion on the basis of the picture very quickly.
Care Mentor AI using the supercomputer of Skoltech will create service for determination of extent of defeat of COVID-19
Within the program for fight against COVID-19 of the Center of Skoltech for scientific and engineering computing technologies for tasks with data bulks (CDISE) date scientists of Care Mentor AI company could use the Zhores supercomputer for increase in accuracy of determination of pathologies. On September 2, 2020 the Care Mentor AI company reported about it. Read more here.
Announcement of Siemens AI-Rad Companion - AI systems for automation of routine tasks in MRT of a brain and prostate
At the end of August, 2020 Siemens Healthineers submitted two new applications for interpretation MPT- pictures on a basis artificial intelligence (AI). New software it is directed to exempting radiologists from routine tasks during MRT of a brain or a prostate gland. In more detail here.
The Botkin.AI platform helps the Moscow doctors to reveal lung cancer on computer tomograms
In Moscow began to analyze computer tomograms and to reveal on them symptoms of cancer of lung artificial intelligence. On August 18, 2020 reported about it in Skolkovo Foundation. The Botkin.AI platform developed by the resident of Skolkovo Foundation Intellodzhik company is for this purpose used. Service is integrated from Uniform radiological by an information system of Moscow. Read more here.
Free AI service of Sberbank will help to reveal changes in lungs, including at COVID-19
The Sberzdorovye service and cloud platform of SberCloud entering an ecosystem of Sberbank started on July 14, 2020 the joint project on recognition of pictures of a computer tomography of lungs. Read more here.
In Moscow earned the first service according to the analysis of roentgenograms using a neuronet
The software of Care Mentor AI which analyzes was integrated into the Uniform Radiological Information Service (URIS) to which the diagnostic equipment of Moscow is connected and carries out screening of roentgenograms of bodies of a thorax on presence of different pathologies, including such socially important as lung cancer, tuberculosis, pneumonia. This Zdrav.Expert became known on July 3, 2020. Read more here.
The producer of KT-scanners released an AI system which makes diagnoses in 10 seconds
At the end of May, 2020 the Chinese producer of computer Imsight Technology tomographs provided an AI system for KT-diagnostics of lungs which sets the preliminary diagnosis in 10 seconds. Development received the name Imsight CT Analysis System. Read more here.
The Russian platform using AI attracted 100 million rubles to diagnosis of cancer
The Intellodzhik company developing the TeleMD Botkin.AI project attracted 100 million rubles from Digital Evolution Ventures funds (it is created with the assistance of Rosatom) and RBV Capital (it is based by Alexey Repik's R-Pharm and the Russian venture capital company) and also the current investors — Primer Capital and "Expokapital". Botkin.AI develops a system based on the artificial intelligence (AI) for the analysis and determination of pathologies in pictures of a computer tomography, X-ray and mammography. The company says that already implemented pilot projects on early diagnosis of cancer of lung in four regions of Russia. Read more here.
The Russian clinic implemented a neuronet of the domestic developer for the description of x-rays
In the Russian private clinic "Meditsina" the neuronet for the description of x-rays developed by Care Mentor AI company (Kerementoreyay) is implemented. It is the unique project for the Russian medicine: instead of 20 minutes which are spent by the doctor for the description of the x-ray the artificial intelligence performs this work in only 3 seconds. So by the time of when the patient puts on after the conducted research, the description of the picture and the preliminary diagnosis will be already ready. Read more here.
AI began to predict capabilities of children according to MRT-pictures
In the middle of March, 2019 the research according to which the artificial intelligence is capable to predict capabilities of children according to MRT-pictures is published. Researchers claim that, analyzing communications of white substance in the child's brain at the birth, the algorithm of AI can predict the level of its cognitive development at the age of 2 years and offer necessary interventions to children from risk group.
Though data demonstrate that the fundamental scheme of a human brain exists already at the birth and that connection of white substance keeps development of functions of a brain, how exactly it occurs – it is unknown. Therefore the team of researchers from Medical faculty of the University of North Carolina led by doctor Jessica Girault decided to train AI in the analysis of communications of white substance for MRT of a brain at the birth at the full-term children.
Using the diffusion weighed MRT-images received soon after the birth researchers allowed AI to distribute 75 full-term babies on points above or below average cognitive development in a scale of ELC used at the age of 2 years. It was found out that AI with an accuracy of 98% predicts correlation between indicators of communications of white substance and the actual cognitive assessment of these children in 2 years. When testing on separate selection of 37 premature children the algorithm showed to 96% accuracy.
Thus the research showed that the network of white substance at the birth has high degree of a predskazatelnost and can be a useful biomarker at visualization. The fact that researchers could reproduce data retrieveds in the second group of children is the convincing proof of reality of their opening. Potentially similar AI can be used for early identification of violations and decision making about remedial measures.
AI for the first time began to predict successfully metastases in MRT-pictures of mammary glands
At the end of January, 2019 researchers reported that they could use rather small data set for training of artificial intelligence in assessment of diagnostic images. According to article published in Journal of Digital Imaging, the research group prepared the program for forecasting of chances of development of metastases in axillary lymph nodes at a breast cancer according to MRT. The AI program reached only 84% of accuracy of diagnostics, but the command is going to collect more data for training an algorithm finally to use it in clinical practice. Read more here.
The artificial intelligence learned to reveal Alzheimer's disease for 6 years before doctors
The artificial intelligence can diagnose Alzheimer's disease of doctors much earlier. Scientists trained a computer system to reveal in pictures of the positron emission tomography (PET) the first symptoms of a serious illness hard to distinguish for eyes. Results of a research were published on November 6, 2018 by the Radiology magazine.
For diagnosis of Alzheimer's disease the team of researchers from the University of California created a self-training AI algorithm. To train a system, scientists used more than 2100 pictures of a brain over 1000 patients who were made using PET-scanning, The Sun tells.
The positron emission tomography allows to estimate metabolic activity of a brain. At Alzheimer's disease metabolic rate in certain sections of nervous tissue decreases, and the artificial intelligence learned to detect these scarcely noticeable changes. 90% of pictures served for training of an algorithm, and other 10% were involved during testing.
During final testing a system managed to reveal in 100% cases symptoms of dementia in pictures of 40 patients, and traced them on average for six years earlier, than the official diagnosis was made.
|We are very happy with work of an algorithm. It identified all patients for whom afterwards diagnosed Alzheimer's disease — one of authors of a research doctor Dzhaye Huo Son (Jae Ho Sohn) said.|
In turn, professor Noel Sharkey from the University of Sheffield, Great Britain, said that though the amount of selection and the program of tests were rather small, the received results are very promising and demonstrate expediency of more large-scale research.
According to specialists, earlier diagnosis of Alzheimer's disease will allow to slow down or even to stop its development.
Facebook by 10 times accelerated MRT-scanning thanks to artificial intelligence
Together with the University of New York the AI system trained approximately for 3 million MRT-pictures of heart, liver and bones and also on 10 thousand researches with the taken-out diagnoses was created. Facebook in this project is responsible for technologies of computer vision and visualization.
Based on a training the neuronet allowed MRT-devices to scan quicker human bodies and to issue results thanks to the fact that to the equipment began to study only some parts of the body enough. Thanks to application of an AI algorithm the speed of operation of medical scanners increased to 10 times.
|Using AI, it is possible to collect less data and, as a result, to accelerate scanning, at the same time having saved or having even expanded MRT-images rich with information. The task consists in training artificial neural networks to distinguish basic structure of images that will fill the spaces arising at the accelerated scanning, said in the blog of Facebook. — This method is similar to how people process touch information. When we learn the world, the brain often receives an incomplete picture as it in a case with indistinct or dimly lit objects, and completes it afterwards. We need to use these data with advantage for medicine.|
Facebook highlights that training of an AI system happened to use of anonymous data in full accordance with regulations of privacy of HIPAA.
Earlier in 2018 the CNBC TV channel reported that Facebook conducts negotiations on a data exchange occasion on patients with hospitals, however the project was closed after its publicity.
Algorithm of Google for the analysis of fluorography of bodies of a thorax
At the end of March, 2018 the Google company showed artificial intelligence for fast and effective processing fluorography of bodies of a thorax. Development was provided at the EmTech Digital conference to San Francisco.
Google created the model of deep training capable to be improved on the basis of a small number annotated (with the areas marked manually on which deviations) medical pictures are visible and the disease and to select it on the image allowing to identify at the same time.
For training of an algorithm used ChestX-ray8 — the largest open database of radiological researches of a thorax which is kept by National Institute of health care of the USA (U.S. National Institutes of Health). More than 110 thousand fluorographic pictures connected with 14 types of diseases and also 880 images for whom the certified radiologists selected 984 areas showing deviations are presented in the directory.
All pictures passed through so-called convolution neural network to sort all information in pictures and to code data, such as disease type, anomaly location, etc.
Then for obtaining local information the image is divided into a set of areas. Thanks to it doctors can quickly reveal illnesses and make diagnoses even without extensive knowledge in the area.
|This algorithm exceeds the modern machine learning used for forecasting of diseases, and, what is more important, issues an analytical picture of the decision made by the computer to help radiologists better to interpret result — the head of department of a research and developments of cloud artificial intelligence and machine learning of Google Jia Li said.|
The artificial intelligence exceeded radiologists in diagnosis of pneumonia
In November, 2017 researchers from Stanford University provided a self-training algorithm (a so-called neuronet) CheXNet which is capable to make the diagnosis to pneumonia according to roentgenograms of lungs. Scientists published results of the work in open access. The received program extremely highly specialized, however copes with the work better than professional radiologists.
CheXNet trained at the public database containing more than 100,000 roentgenograms of a thorax on which it is possible to distinguish 14 pathologies. After training the neuronet was checked: to several radiologists suggested to carry out the analysis of test x-rays, and results compared with diagnoses of the machine. As it appeared, the computer system could diagnose pneumonia more precisely, than the person.
Pneumonia — a dangerous and widespread disease, and its early identification will help to prevent a set of death; only in the USA about 50,000 people annually perish from pneumonia. Besides, pneumonia is one of the main reasons of child mortality from infectious diseases.
Developers taught to mark a system in different flowers those departments of lungs where the machine "saw" symptoms of pneumonia; the color is brighter – the pathology is more probable. And after machining of the roentgenogram the doctor browses, paying attention first of all to those areas which the machine marked as "hottest".
Andrew Ng, the coauthor of article and the former head of group of a research in the field of artificial intelligence in Baidu companies, believes that the similar systems soon will begin to be applied everywhere. Geoffrey Hinton, one of pioneers of development of systems of deep training, considers that need for training of new radiologists disappeared, and the neuronet can quite cope with their functions. Except pneumonia, computer systems are also able to detect signs of presence of tumors, violations of a rhythm of heart and other pathological changes on x-rays, electrocardiograms and other systems of visualization.
The overview of development of AI in radiology
At the beginning of May, 2017 the AuntMinnie.com edition published article devoted to use of the artificial intelligence (AI) in radiology. The publication contains an output that the computer is not ready to replace radiologists yet, however is capable to influence considerably their activity, having optimized workflows, having reduced the number of inspections and having even rendered the help in search of genomic markers.
As doctor Eliot Siegel from the University of Maryland reported, for the last 30 years thousands of works devoted to analysis algorithms of medical images however are published the few from them found application in medical practice. Those technologies which nevertheless were adopted are rather evolutionary, but not revolutionary, he noted.
|Mammographic detection systems with computer data processing exist more than 20 years and in fact are the only programs of machine learning which are widely applied in diagnostic imaging today. About 90% of mammographs use it, however with a huge share of doubt and scepticism, and in certain cases and with a sneer — Siegel reported.|
However, according to him, use of graphic processors for machine learning led to significant increase in computing power. It facilitated implementation of the most exacting to resources of computing methods, such as deep training, image understanding and convolution neural networks.
All this has high potential for work with medical pictures, however requires careful marking and drawing up detailed descriptions of images manually and also a large number of researches on visualization, the expert added.
Doctor Marc Kohli from the University of California in San Francisco agrees that the artificial intelligence is still far from clinical practice of the average radiologist, despite numerous projects in Silicon Valley.
At the same time more and more radiodiagnosis specialists gain basic knowledge about machine learning and their application in practice. Academicians hope that the growing interest in artificial intelligence will increase investments in this area. At the same time processing and data analysis specialists even more often enter many academic programs, notes Kolya.
Doctor Bradley Erickson from Mayo clinic in Rochester (the State of Minnesota, the USA) says that by 2017 the most widespread field of use of machine learning is speech recognition, however there are a lot more directions for improvement of technology.
According to doctor Luciano Prevedello from the medical center of Ohio State University in Veksner, development of artificial intelligence in radiology proceeds quickly, however this sphere still is in embryo.
Prevedello expects that solution AI will begin to be used more often by 2019 for optimization of workflows. For example, in one of Ohio clinics "smart" algorithms for search of pictures of a computer tomography with critical indications are already used. Use of AI for setting of diagnoses according to medical pictures will be perhaps later because of need of carrying out numerous tests before implementation, the doctor considers.
Doctor Raymond Geis from Medical school of the Colorado University considers that AI is useful to search of patterns in data and forecasting of behavior on the basis of the previous models, and the first commercial AI algorithms in radiology will not analyze images. More likely, they will be used for drawing up reports on the basis of outputs of doctors.
Commercial solutions can be used, for example, in KT-scanning for warning of doctors of dangerous deviations, takik as sharp intra cranial hemorrhage, assumes Geys.
Eliot Siegel sees big perspectives of artificial intelligence in specialized areas: determination of fractures of bodies of vertebras, detection and drawing up characteristics of nodes in lungs, MRT and the nuclear analysis of images of heart.
Mark Kolya found it difficult to call what areas will be the most important for clinical use of artificial intelligence in radiology, but expressed opinion that success of projects will depend on many factors, including on correctness of integration and checks and also on solutions of regulators.
Applications which do not need approvals of regulating authorities will quicker develop as the authorities have no debugged mechanism allowing PACS developers and other similar systems to offer AI software created by third-party developers because of strict requirements to writing of the code, documentation and testing.
One more problem, according to Ziegel, are high system requirements to computers on which algorithms of machine learning are executed. The solution of this problem there could be cloud services, however many radiologists do not trust them and not actively use them, he noted.
Bradley Eriksson is sure that the artificial intelligence, first of all, will be useful to display of genomic or diagnostic properties which radiologists do not see today.
|We and others publish the researches showing that deep training can predict genomic markers with a high accuracy according to normal images of KT and MPT even if people it is not enough or at all nothing is known about these signs" — Eriksson told.|
Spanish software with AI for recognition of a disease according to x-rays
Within the congress of the European society of radiologists (ECR-2017) which passed in Vienna from March 1 to March 5 the Spanish Medical Research Institute La Fe submitted the software using artificial intelligence for primary detection of diseases according to x-rays. Read more here.
You See Also
- Telemedicine (Russian market)
- Telemedicine (world market)
- Remote monitoring of health of patients
- ↑ AI can predict cognitive development from MRI at birth
- ↑ Artificial intelligence can predict Alzheimer’s six years earlier than medics, study finds
- ↑ A Deep Learning Model to Predict a Diagnosis of Alzheimer Disease by Using 18F-FDG PET of the Brain
- ↑ Facebook and NYU School of Medicine launch research collaboration to improve MRI
- ↑ Google AI algorithm shows promise for chest X-rays
- ↑ Can AI diagnose pneumonia better than radiologists?
- ↑ Part 2: How will AI affect radiology?
- ↑ Part 1: How will AI affect radiology?