Semantic Segmentation of Mammograms Using Pre-Trained Deep Neural Networks
Rodrigo Leite Prates, Wilfrido Gómez‐Flores, Wagner Coelho de Albuquerque Pereira · 2021 18th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE) · 2021
Anatomic regions like breast and pectoral muscle are common regions that need to be segmented before determining abnormalities in the mammographic image. For this task, we explore five convolutional neural network models, namely, Resnet50-Unet, Mobilenet-Unet, Vgg-Unet, Unet, and Segnet. A classic technique that uses an MLP network is compared with the convolutional methods. The MIAS and INbreast datasets are used for evaluating Deep Neural Networks on the segmentation of these regions. Several evaluation metrics are used to compare the results of convolutional approach and a traditional method based on hand-crafted features. The results confirm the superiority of the deep networks over the traditional method. The best Jaccard value for Deep Learning algorithms is$0.945\pm 0.068$. The Mobilenet-Unet network presents the best trade-off between computational cost and segmentation capacity; therefore, it is the best choice to compose a CAD system.