Segmentação automática de lesões de mama em imagens de ultrassom utilizando redes neurais convolutivas

Abstract

Early diagnosis is a crucial factor in increasing the chances of breast cancer treatment. Mammography is currently the best way to detect lesions in the early stages, however, in some cases, it is inconclusive, requiring ancillary exams to obtain a differential diagnosis. Given this scenario, breast ultrasound appears as the main adjunct because of its cost and accessibility. Breast ultrasound can differentiate malignant lesions from benign through features such as shape and contour, however, its analysis is not a trivial task, which can become very costly and contain variability. Because of this, computational methods have been created to assist the experts in this task. This dissertation develops a computational method, based on convolution neural networks, to segment breast lesions in ultrasound images. Three architectures (BUS-CNN1, BUS-CNN2, BUS-CNN3) are developed with different topologies in order to analyze the best architecture for this task. The database used contained 387 breast ultrasound images and was divided into training and test sets with 255 and 132 images, respectively. Six performance metrics for quantitative analysis were used: accuracy, global accuracy, Intersection Over Union (IOU), weighted IOU, Boundary F1 (BF) ratio and Dice coefficient of similarity. The three architectures were trained and tested with the same sets. After the tests, it was shown that the BUS-CNN3 architecture obtained the best results in five of the six metrics used, with a global accuracy of 95.93%, IOU of 87.92%, weighted IOU of 92.36%, BF ratio of 68.77% and Dice coefficient of 89.11%.