Development and Research of Segmentation and Object Recognition Algorithms on Medical Images Based on Neural Networks

Abstract

This paper discusses the application of neural network architectures for segmentation and object recognition 
tasks on medical images. The base models used are U-Net, Attention U-Net, DeepLabv3+, and EfficientNet, tested 
on open medical datasets BraTS, ISIC, and COVIDx. The results of training, visualization, and comparative analysis 
are presented, and limitations and directions for improvement are discussed, including the use of multimodal data and 
self-supervised learning. Experimental comparison of models is conducted based on the Dice, IoU, Accuracy, and F1
score metrics.