ITMO: Method for neural network model recognition of brain tumors from MRI images

Main articles:

• Neural networks (neural networks)
• Artificial Intelligence in Medicine
• Artificial intelligence in radiology

2021: Creation of a method for automatic recognition of brain tumors

Graduate student of the faculty information technology programming ITMO and created a method for automatic recognition of tumors from brain images. It MRI will allow better interpretation of predictions neuronets to determine where she is unsure. This was announced by the University on August 26, 2021.

The method is based on deep learning methods, it aims to reduce the level of uncertainty of the neural network model when analyzing medical images. The fact is that doctors are engaged in marking medical images, manually performing a huge amount of labor-intensive work. But this is fraught with errors and inaccuracies, which means it can affect the work of the neural network model. At the same time, the problem with neural networks is that a person fully trusts the decisions of the model. But you cannot blindly rely on her conclusions, instead you need to take into account how confident the model is in its prediction. This is especially important for solving problems in the field of medicine.

Compared to basic neural networks for processing medical images, the accuracy of the ITMO researchers algorithm is 3% higher, and the calibration of the model has become twice as good. In fact, this is quite a lot - such an algorithm will not only be able to determine the pixels representing the tumor, but also predict its boundaries much better, and also more accurately tell the doctor in what pixels of prediction he is sure of more or less.

The algorithm we created solves the problem of tumor segmentation. Suppose there is an image of the brain, the neural network will receive it and convert it into a binary picture with pixels marked 0 or 1, each of which will correspond to a healthy tissue site or neoplasm. The algorithm also allows you to see areas in which the machine learning model is less confident, which require closer attention of the doctor. These are usually tumor boundaries. Our model is better calibrated because it has a much lower percentage of uncertainty, "said Natalya Khanzhina, author of the project, graduate student of the ITMO Faculty of Information Technology and Programming.

The algorithm was tested on an open set of data BraTS, including 45 thousand images of magnetic resonance imaging. This is very valuable for the scientific community, since the method is available, universal, and, therefore, can be used to solve other problems.

Natalia is working on the study with Maxim Kashirin, a graduate of the Machine learning and Data Analysis master's program at ITMO University.