Neural network models for analyzing MRI images of pituitary adenoma

Abstract

The purpose of this work is to study methods for intelligent analysis of MRI images of the sellar region for diagnosing pituitary diseases. To solve this problem, it is proposed to create a deep learning neural network architecture based on the PyTorch framework. To train, test and validate the resulting architecture, data from digitized MRI images of the brain with varying degrees of invasion, а contrast-enhanced T1-weighted (T1-CE), a T2-weighted (T2) in the amount of 282 units, obtained from the National Medical Research Center of Endocrinology of the Ministry of Health of Russia (NMRC of Endocrinology) were used. Using the data obtained, comparisons were made of the proposed neural network architecture with known implementations of machine learning architectures when solving problems of classification according to the international Knosp scale and segmentation of pituitary adenomas. Recall was chosen as the target metric for assessing the quality of classification, and Intersection over Union (IoU) was chosen for segmentation. The results obtained confirmed the effectiveness of the proposed solution. The result of the work was the creation of an effective architecture for a machine learning system used for classification and segmentation of pituitary adenomas.