From an excess of data
How the diagnosis of dangerous diseases is put on stream
RIA Novosti, Tatiana Pichugina.
Billions of images obtained by various methods are stored in medical databases around the world. These visual materials are ideal for training neural networks and automating diagnostic processes. About the successes of Russian scientists in this direction — in the material of RIA Novosti.
AI is walking on the planet
The idea of artificial intelligence (AI) was formulated in the middle of the XX century, but the breakthrough occurred only with the advent of modern computing power. Now AI is beating a person at chess, controls drones, recognizes faces and voices, improves photos and solves many more tasks that are often invisible to us.
AI has great prospects in medicine: new diagnostic equipment and, as a result, an avalanche of various data contribute to this. The volume of visual content, and with it the burden on diagnostic doctors, is growing exponentially. Since the 1980s, scientists have been creating automated complexes for processing radiological images. However, clinical studies have shown that programs at best do not help radiologists, at worst they complicate their task due to the high percentage of false positive diagnoses.
Progress has been made in recent years — thanks to a multiple increase in the power of servers and GPUs. The new hardware gives an impetus to the development of deep learning neural networks. In 2012, the term radiomics appeared - the science of working with large visual data in medicine. Traditional AI analyzes images based on signs known to researchers. The method of deep machine learning consists in the fact that the neural network itself extracts features from an array of information and offers a solution to the problem.
To help radiologists
Nikolai Staroverov and his colleagues began to implement the idea of analyzing X-rays in 2017, while still at the senior courses of LETI in St. Petersburg. They wanted to teach AI to identify lung pathologies. Staroverov received an RFBR grant for this task, for which he purchased computer equipment. Software tools, for the most part, were in the public domain, in libraries for creating neural networks. Scientists take ready-made solutions there, refine and embed them into their own architecture (it is this architecture that largely serves as know-how). Medical data was also used openly or obtained through familiar doctors. At first, there was not enough information.
After several presentations at conferences, Staroverov's project was noticed by the Moscow company Med-ray, which creates automation systems for storing and processing medical images. The entrepreneurs decided to invest in the research of young scientists in order to get a domestic module in the future that can be implemented into their system. And they provided 1.6 thousand X-ray images of the lungs, already marked by doctors. On this array, neural networks were trained to detect pathologies: pneumonia, atelectasis, hydrothorax, focal shadows in the lungs, even rib fractures.
"The training is quite long, but already a year ago we achieved good results. The program determines the norm in 97 percent of cases. With pathologies, it is worse — the accuracy so far reaches 75 percent, in the case of emphysema of the lungs — up to 90. It is necessary to improve the methods of image preprocessing," says Nikolai Staroverov, assistant of the Department of Electronic Devices and Devices at SPBSET "LETI".
It is necessary to clean the images from extraneous noise — for example, numerous wires that entangle a patient in intensive care can get there. There is also a difficulty with verifying the initial data, the scientist explains, because the doctor himself sometimes makes the wrong diagnosis.
Technically, everything looks simple: the diagnostician opens a snapshot on the screen, turns on the AI option and, if a pathology is detected, looks more closely. This will help when working with a large number of images.
"During the day, the radiologist has more than fifty patients, gradually he loses concentration. And the program will make an initial diagnosis for him, thereby taking on part of the work," says the scientist.
During the COVID-19 pandemic, a lot of lung scans appeared. Many took up their analysis, competition arose. "Our advantage is that we identify several different pathologies. The next step is to identify malignant tumors," explains Staroverov. He cites the example of focal shadows in the images, which indicate either tuberculosis lesions or oncology.
By the summer, the developers hope to achieve the necessary accuracy in analyzing the images, after which they can start creating a commercial product. But this is a matter for investors, and researchers will deal with other tasks.
Functional map of the brain
Skoltech works with brain images and signals obtained by various methods, such as electroencephalography, magnetic resonance imaging (MRI), functional magnetic resonance imaging — fMRI, functional near infrared spectroscopy (fNIRS), eye tracking (by eye movement), intracranial angiography, magnetic encephalography. First, the data needs to be cleaned from noise and external induced fields, then it needs to be prepared for further analysis by deep neural networks. These are programs that themselves identify a set of features in an array and pass them on for analysis.
"The classic task is classification, when it is required to distinguish pathology from the norm. Another approach is quantitative, when it is necessary to predict what the patient will have with speech, for example, in a year. Neural networks can map different functional parts of the brain, as well as make descriptions of images," says Candidate of Physical and Mathematical Sciences biophysicist Maxim Sharaev, senior researcher at the Center for Artificial Intelligence Research.
Scientists are working with the Burdenko Neurosurgery Center on a system that automates one of the important preparatory stages for brain surgery.
"David Pithelauri, head of the Department of glial Tumors, was the first to show interest. He is a world—renowned scientist, he knows how to set tasks and assess their relevance in medicine," Sharaev notes.
We are talking about mapping the cerebral cortex according to fMRI data. This will help the surgeon to determine more precisely which area of the skull to open, and during the operation he will see which departments cannot be affected so as not to damage vital functions. Now this is being found out experimentally. The doctor touches the brain with an electrode and finds, for example, the motor area responsible for movement, or speech — the patient is conscious and answers questions. Operations usually last many hours, sometimes a whole day. A preoperative brain map will significantly reduce the time.
"Already now the surgeon can compare two images on the monitor — the original and the mapped one. Ideally, we want to create a neuronavigation to see the projection of departments on the brain right during the operation," says the biophysicist.
In search of a minor pathology
Scientists use the same approach to solve another problem — the search for small pathologies that cause epilepsy. We are talking about millimeter-sized areas of gray matter with impaired organization of neurons, due to which children develop epileptic seizures. This form of the disease is called focal — in contrast to generalized, characterized by abnormal activity in the entire volume of the brain. If you remove this center of epilepsy, the disease passes: the child catches up with peers in development, the adult no longer needs medication. Operations are successful in 90 percent of cases with the correct diagnosis.
There are not enough specialists in the diagnosis of focal epilepsy in Russia, and the analysis itself takes a lot of time, because a doctor should notice a very small area with blurred borders on a lot of images. Sometimes a consultation is required.
How difficult it is to diagnose focal epilepsy is shown in these images. On the left — the white spot of the tumor is clearly visible, on the right — focal cortical dysplasia, which can only be distinguished by a specialist. Photo here and below: Maxim Sharaev/Skoltech.
There are several automated solutions to similar problems in the world, and two are certified by the American FDA. But their capabilities are limited, there are disadvantages. Therefore, scientists from Skoltech decided to create their own model. The data and expertise were provided to them by specialists from the Scientific Center of Obstetrics, Gynecology and Perinatology named after Academician V. I. Kulakov.
"In two years, we have collected a high—quality set of data - images of more than 200 patients with confirmed diagnoses. Each image is marked up by radiologists. We have trained the program to outline approximate areas of the brain to search for pathologies. The doctor analyzes them and makes a diagnosis himself," explains Maxim Sharaev.
Scientists use ultra-precise neural networks (3D CNN) to work with three-dimensional images (technically, this is an array of two-dimensional head shots). Another approach borrowed from self—driving cars is point clouds, when the area is represented not by a dense, but by a sparse set of values. This significantly saves RAM and computing resources of the video card on which the neural network is trained.
Identification of abnormal areas of the brain associated with epilepsy using a neural network.
"There are many open libraries, from where we take the blanks, assemble the architecture of the neural network, and then train it. The more complicated it is, the more time is needed. Each block of the model has a huge number of parameters that need to be selected. We submit images to the input — we get predictions at the output. It is not necessary that they are very accurate, it is enough to show an approximate search area. The main thing for a doctor is not to miss pathology with such a hint," the scientist says.
Together with colleagues on a grant from the Skolkovo Foundation, specialists from Skoltech have developed a platform with a user-friendly interface that allows you to store all the necessary data, mark them up, show a map of the probability of finding a pathology. The doctor can move it to get a better look at the anomalies.
Now a prototype of the system is ready, after completion, a startup will grow out of it. The authors will register a patent, test the invention in real conditions and enter the markets.
Portal "Eternal youth" http://vechnayamolodost.ru