Comparative analysis of modifications of U-Net neuronal network architectures in medical image segmentation

Data processing methods based on neural networks are becoming increasingly popular in medical diagnostics. They are most commonly used to evaluate medical images of human organs using computed tomography, magnetic resonance imaging, ultrasound, and other non invasive diagnostic methods. Disease diagnosis involves solving the problem of medical image segmentation, i.e. finding groups (regions) of pixels that characterize specific objects in the image. The U-Net neural network architecture developed in 2015 is one of the most successful tools to solve this issue. This review evaluated various modifications of the classic U-net architecture. The papers considered were divided into several key categories, such as modifications of the encoder and decoder; use of attention blocks; combination with elements of other architectures; methods for introducing additional attributes; transfer learning; and approaches for processing small sets of real world data. Different training sets with the best parameters found in the literature were evaluated (Dice similarity score; Intersection over Union; overall accuracy, etc.). A summary table was developed showing types of images evaluated and abnormalities detected. Promising directions for further modifications to improve the quality of the segmentation are identified. The results can be used to detect diseases, especially cancer. Intelligent medical assistants can implement the presented algorithms.