Magnetic resonance imaging (MRI) is the most common method for primary non-invasive diagnosis of brain tumors and monitoring the dynamics of the disease. Some of the most challenging tasks in this field are classifying tumor types and determining their clinical parameters such as size and volume. This data is important for diagnostics and treatment procedures, including surgical operations. Researchers fr om the NSU Mechanics and Mathematics Department Laboratory of Streaming Data Analytics and Machine Learning have developed a software module for the differential diagnosis of brain tumors on MRI images. The algorithms were developed with the support of a grant from the Russian Foundation for Basic Research (RFBR), and the software module was developed within the framework of the Priority 2030 program.
Evgeniy Pavlovsky, Laboratory Head, described their work,
We integrated the created module into freely distributed open software that is well known among radiologists. This software allows technicians to view 3D MRI images and is very easy to use. We built a module into it in which we packaged our algorithm. It is quite complex and heavy so it is executed on a powerful remote server and the module is the client element. The module sends MRI images to the server and receives the result of tumor segmentation and classification as a response. Our development is based on two- and three-dimensional computer vision models with pre-processing of MRI sequence data. These models make it possible to detect and recognize four types of tumors with high accuracy (meningioma, neuroma, glioblastoma, and astrocytoma) as well as to identify their components and sizes.
The software module is based on a data set developed by Laboratory researchers, which contains information about more than a thousand neurosurgical patients with postoperative diagnoses confirmed by histological and immunohistochemical methods. This data set was generated jointly with the Federal Center for Neurosurgery (Novosibirsk).
Pavlovsky continued.
We supplied this database with expert markings, which were carried out by radiologists from the Federal Center for Neurosurgery and the Research Institute of Clinical and Experimental Lymphology, a branch of the Federal Research Center Institute of Cytology and Genetics SB RAS. The Head of the Institute of Cytology and Genetics SB RAS, Andrei Letyagin, contributed his expertise to the tumor marking by confirming or rejecting it. - Representatives of the Federal Center for Neurosurgery carried out similar work. On this basis, we trained artificial intelligence algorithms. They, of course, do not yet have the property of explainability, but our next step will be devoted to this problem. We plan to teach the algorithm to explain why it produced the contour of the tumor and classified it exactly like that. This will not be easy to do, but the existing algorithm has a high degree of confirmation among experts and produces information with a high degree of objectivity. We have proven this in our scientific articles published in several specialized publications. And, perhaps, in the future our algorithm will replace a person in performing routine processes determining the contours and types of brain tumors on MRI images.
The researchers tested the algorithm on their own database and were satisfied with the results. Then testing and verification took place at the international competition BraTS Challenge 2021. The accuracy of the algorithm was carried out on a test data set using the Dice metric. As a result, the algorithm developed by NSU specialists entered the top ten, ahead of almost a thousand teams.
Pavlovsky explained their results,
According to the classification, the results obtained demonstrate not only high accuracy rates of up to 92% when analyzing patients, but also very high potential for future research in this area. The software module we developed can be used to train specialists. It鈥檚 too early to talk about its introduction into clinical practice because it has not yet passed the tests necessary for this.
The Laboratory specialists are currently on another project that is also related to the study of MRI of the brain for other diagnoses, specifically multiple sclerosis. Here they face serious challenges because the pathology foci are much smaller in size than tumors and only outlining contours is not enough. If the tumor grows slowly in one place, then lesions of multiple sclerosis appear and disappear. Artificial intelligence will need to not only record them at a specific moment, but track wh ere they moved or how they united with each other.
Pavlovsky concluded,
Scientists at our Laboratory see its future development in the application of artificial intelligence methods to medical problems. We intend to continue working on this and to train Master鈥檚 and Graduate students capable of developing it.
