Russian neural network provides a second opinion for diagnosing vertebral problems
The diagnostic system, developed by Nanosemantics (a Skolkovo resident), together with the American IT startup Remedy Logic, is capable of detecting central, lateral and foraminal stenosis of the lumbar spine, arthrosis and hernia on MRI images. Artificial intelligence not only detects pathology with high accuracy, but also offers treatment options.
It all started when Remedy Logic noticed that many expensive spine surgeries are performed in the United States every year, which could have been avoided if the analysis of MRI scans had been more accurate. They decided to involve artificial intelligence in diagnostics and turned to colleagues from Nanosemantics to create software for processing medical images. The Russian company has developed the necessary algorithms.
Each MRI image passes through two neural networks: a classifier and a segmenter. A classifier is an algorithm that assigns each object to one or another predetermined class, based on the characteristics of the object. Segmentator is an algorithm for selecting objects and their boundaries in the image. For the segmentation of physiological objects in the images, a publicly available neural network of the U-Net architecture is used, adapted to the goals of the project. For qualitative analysis of images, a classifier based on the MobileNetV2 architecture is used. It has high accuracy and high speed (it can be run even on mobile devices).
MRI images for training artificial intelligence “Nanosemantics” were provided by the Federal Center for Traumatology, Orthopedics and Endoprosthetics of the Ministry of Health of Russia (FTSTOE) in Cheboksary and the network of clinics “MRT Expert”. There, programmers received expert advice from traumatologists and radiologists, as well as statistical data on the condition of patients after treatment.
The system was launched on the Remedy Logic website. To begin with, the patient answers the questions of the questionnaire – it was prepared by the American partners. Questioning is an important part of the diagnosis. “If we were just processing the images, the diagnosis would be less accurate,” explains Vladislav Tumko, lead developer of Data Science at Nanosemantics, project manager for creating a diagnostic system. “Let's say our system sees a large hernia in the image, but the patient does not have severe pain feels, moves normally. Without a questionnaire, the system would not take into account the patient's feelings and would suggest an operation, although in reality physical therapy or medication could be dispensed with. “
After filling out the questionnaire, the user uploads the MRI images to the system. The system analyzes them and issues a detailed diagnosis along with treatment recommendations based on statistics of similar cases. “On a sample of two or three patients, it is difficult to see the overall statistical picture,” says Vladislav Tumko. “If we take 500 thousand patients with similar pathology, then we can identify certain patterns. Let's say that out of these 500 thousand people, someone took medications, someone underwent physiotherapy, someone went to the operating table. After analyzing all these cases, we find out that, for example, physiotherapy gave the most positive results, and, suppose, during operations, there were no more than 20% successful cases, and 30% had complications, the remaining 50% did not receive the expected positive result. We enter all this information into our system. Having received the patient's data, the system calculates everything and offers him, for example, to undergo physiotherapy, while showing other options. It does not impose anything, leaving the patient the right to choose himself. “
The system provides high diagnostic accuracy. “When a radiologist takes pictures, he must view all areas of the lower back and all this in different projections,” explains Vladislav Tumko. “He has hundreds of images of one patient alone. And if there are many pathologies, he may miss something in the process. The system, on the other hand, does not focus on any one area, but completely examines the entire image, and all areas are equally attentive, and, unlike a doctor, cannot miss anything. She conducts analysis and all calculations several times faster and more accurately than a human. “
And here's another interesting thing, the project leader notes: in the course of a study that the company conducted, comparing the work of doctors with the work of artificial intelligence, each doctor interpreted what he saw in the pictures in his own way. One believed that there was a pathology, the other – that everything was within the normal range. In assessing the severity of the condition, they also differed: one was sure that the condition was of moderate severity, the other insisted that the situation was critical and required surgical intervention. A system trained on a large number of opinions is much more objective. “It eliminates mistakes not only due to fatigue, but also due to the difference in experience,” Tumko emphasizes.
Now the specialists of Nanosemantics are working on expanding the functionality. In the near future, they will learn to identify all the pathologies of the lower back, then they will move on to the cervical and thoracic spine, and then to the limbs. The company plans to Russify the diagnostic system and install it in the MRT Expert network of clinics within six months to help radiologists.