Evaluating the Performance of Convolutional Neural Networks with Direct and Sequential Acyclic Graph Architectures in Automatic Segmentation of Breast Lesions in Ultrasound Images

Global health rates show that breast cancer has remained at the top of the biggest causes of death among women. Radiologists use images such as ultrasonography to aid diagnosis. In the era of big data, in which there is a large amount of data available and the use of artificial intelligence is omnipresent in assisting activities, automatic diagnosis aid is a topic on the agenda. Convolutional neural networks are efficient in the most medical tasks. In this work, the performance of two convolutional neural networks, one with sequential architecture and the other with a direct acyclic graph structure, are contrasted for the task of automatic segmentation of breast lesions in ultrasound images. For the development and evaluation of the proposals, two image banks were used, containing a total of 550 ultrasound images. Performance metrics already established in the literature, such as Global Accuracy and Dice Coefficient, were used to evaluate the network architectures. The best segmentation result shows a global accuracy of 96%.