The development of neural networks that, having learned from labeled medical data, help doctors make diagnoses and even develop treatment recommendations is one of the most advanced areas for using artificial intelligence technologies in medicine. A project of this kind was recently implemented by students of the MEPhI.
The project is called "Intellectual assistant to the doctor for ultrasound diagnosis of thyroid nodules." As the name implies, the intelligent system helps endocrinologists by classifying neoplasms in the human thyroid gland based on ultrasound images. "Assistant" classifies nodules based on TI-RADS (Thyroid Imaging Reporting and Data System) - an international standardized system for describing and processing thyroid imaging data.
The system was developed by students of the MEPhI with the participation of professor of the Institute of Cyber Intelligence Systems (ICIS) of the MEPhI Konstantin Zaitsev and the teacher of the Higher School of Economics Maxim Dunaev under the general supervision of the director of the Institute of Physics and Technology Alexander Garmash.
The project partner was the National Research Center for Endocrinology of the Ministry of Health of Russia, whose specialists not only advised the developers, but, more importantly, provided labeled medical data for neural network training. In total, data from 137 patients were used, the diagnostic results of which were included in the training sample of 400 unique cine loops (sets of several dozen images) and single images. Before transferring to the developers, all data was anonymized.
From a technical point of view, the “Assistant” was based on the architectures of neural networks already available in the public domain – the Deeplabv3+ neural network architecture was used to segment images and highlight nodes, and the EfficientNetB6 neural network was used to diagnose pathologies and classify neoplasms. However, these programs were taken only as a basis - the members of the student team had to work on the architecture of the final product.
In addition to the actual diagnostic functions, the "Assistant" performs other functions that are needed at the doctor's workplace - these are patient records, medical statistics, and the "expert mail" block - a special interface that helps the doctor get a "second opinion" from another specialist.
It is also very important that the "Assistant" has a built-in feedback system that allows the doctor working with him to point out errors in the neural network, and thus increase its accuracy.
The system has been tested since the beginning of this year on the basis of the National Medical Research Center of Endocrinology, and three practicing doctors are currently working with it. Work is also underway to register "Assistant" as a medical device.
Meanwhile, as the project participant, student Ksenia Tsiguleva, said, the Assistant development team has big plans: in the future, the neural network should analyze data not only from ultrasound, but also from cytological, histological and genetic studies.
In addition, the development team plans to join another project being implemented at the MEPhI - the development of a medical dermatological neural network.
Recently, the “Assistant”, created at the MEPhI, was awarded the International University Prize in the field of artificial intelligence and big data “Gravity – 2023” in the “Breakthrough Research and Development” nomination.
It should be noted that most of the undergraduate students who worked on this project are going to enter the Higher Engineering School of the MEPhI - a master's program that trains system engineers, digital designers and programmers specializing in managing the life cycle of complex technological systems.
