Avaliando o desempenho de redes neurais convolucionais com arquiteturas de grafos acíclicos diretos e sequenciais na segmentação automática de lesões mamárias

Abstract

Breast cancer can be seen as a worldwide problem, which is responsible for a substantial number of deaths. Diagnosis through image analysis of the lesion is efficient, notably through the use of machine learning techniques. The success achieved in recent years has grown due to the use of convolutional networks. This method is capable of successfully performing computer vision tasks, such as the automatic segmentation of lesions in the most varied modalities of biomedical images. This stage, segmentation, supports later stages of a computer-aided diagnostic imaging system. In this dissertation, the performance of convolutional neural networks with direct acyclic graph architectures in the automatic segmentation of breast lesions in ultrasound images is evaluated. Four convolutional network architectures were implemented and tested, three of them with direct acyclic graphs (DAG) and one sequential. For the development and evaluation of the proposals, two banks of breast ultrasound images (bank A and bank B) were used. Some striking differences between these banks are the size, resolution, and quality of the images. Thus, the images were previously processed and adapted (cropping and resize). The training of these networks with a stop for the number of seasons proved to be unstable. This problem was overcome with the proposal of a training aid function, which allowed us to obtain the best performance point of the model. The best architecture was chosen based on metrics already established in the literature, such as global accuracy and the Dice coefficient. The four architectures tested achieved similar results, with a global accuracy of more than 94% each. The t-student statistical significance test showed that for both databases the best network architecture in the validation was U-net, reaching over 99% of global accuracy for database B and over 96% for the database A. The network with the best performance could be tested, with other input data, and its performance remained the same. It was possible to conclude that the cropping procedure was not crucial for good segmentation accuracy. Besides, the statistical analysis of the performance metrics showed that the use of better resolution images (bank B) did not cause statistically significant performance differences (p <0.05).