Comparação de modelos de redes neurais na segmentação de vasos sanguíneos em imagens médicas

Abstract

The analysis of blood vessels is responsible for extracting various important pieces of information in the healthcare field, and the impact of more precise analyses on disease studies and diagnoses holds positive potential. However, performing these analyses manually involves significant time and resource consumption. The segmentation of blood vessels in images represents a large part of this difficulty and cost. Technological advancements have enabled the implementation of machine learning techniques to carry out this crucial task, particularly through the use of neural networks, which has marked a significant evolution in the field. This work analyzes and compares various neural network models that have recently garnered significant attention, such as ResNet and EfficientNet, for example. The goal of this analysis and comparison is to obtain important insights into the advantages and disadvantages of applying each investigated model, such as situations where the models have lower accuracy and the influence of preprocessing and training stages on the models. The impact of varying the size and number of parameters in the models is also investigated, as smaller models may deliver very satisfactory results while consuming fewer computational resources and much less time. Finally, the tests indicated a good result with the models implemented with U-Net, especially RegNet and DenseNet. Also noteworthy were the results obtained by smaller models, like RegNetY_002, for instance, which in several cases achieved similar results in IoU and clDice as models with approximately five times more parameters.