COMPARATIVE ANALYSIS OF MODIFICATIONS OF U-NET NEURAL NETWORK ARCHITECTURES IN THE PROBLEM OF MEDICAL IMAGE SEGMENTATION

Data processing methods using neural networks are gaining increasing popularity in a variety of medical diagnostic problems. Most often, such methods are used in the study of medical images of human organs using CT scan and magnetic resonance imaging, ultrasound and other non-invasive research methods. Diagnosing pathology in this case is the problem of segmenting a medical image, that is, searching for groups (regions) of pixels that characterize certain objects in them. One of the most successful methods for solving this problem is the U-Net neural network architecture developed in 2015. This review examines various modifications of the classic U-Net architecture. The reviewed papers are divided into several key areas: modifications of the encoder and decoder, the use of attention blocks, combination with elements of other architectures, methods for introducing additional features, transfer learning and approaches for processing small sets of real data. Various training sets are considered, for which the best values of various metrics achieved in the literature are given (similarity coefficient Dice, intersection over union IoU, overall accuracy and some others). A summary table is provided indicating the types of images analyzed and the pathologies detected on them. Promising directions for further modifications to improve the quality of solving segmentation problems are outlined. This review can be useful for determining a set of tools for identifying various diseases, primarily cancers. The presented algorithms can be a basis of professional intelligent medical assistants.