Performance of the SegNet in the Segmentation of Breast Ultrasound Lesions
Pedro Vianna, Ricardo Farias, Wagner Coelho de Albuquerque Pereira · 2021
This work presents a performance analysis of a convolutional neural network (CNN), named SegNet, applied for automatic segmentation of breast ultrasound images. Defining the viability of CNN's could help with feature extraction and posterior classification of lesions, causing a decrease in the number of unnecessary biopsies. The breast ultrasound (BUS) dataset consists of 2054 images from 659 female patients, and each image was manually segmented by an experienced radiologist. The dataset is fed into the CNN without any filtering process applied to it and the test output is then compared with the radiologist's labels. We compared the results of different loss and activation functions, and the experiments reveal that the best performance was achieved using dice loss function, with dice coefficient of 81.1%.