ABSTRACT
Computer-assisted diagnosis of breast mass detection and segmentation is becoming more and more of a necessity given the exponential growth of mammograms performed each year. However, the identification of masses on mammography images is not a trivial task, especially for dense breasts. During the last decade, several works have contributed to the improvement of the mammographic masses segmentation; nevertheless, the emergence of deep learning (DL) caused a huge success in image processing due to its performance and its ability to generalize across diverse data. All these observations put scientists and researchers to wonder whether traditional learning methods are still relevant in the segmentation of this type of biomedical images. In this chapter, a comparative study of different segmentation techniques is carried out on INBREAST database, between DL approaches with the Convolutional Neural Networks and a traditional machine learning technique by the super-pixel-based segmentation. The experimental segmentation results of the most common DL networks, namely, Unet, FPN and LinkNet, show that the Unet network outperforms on the task of breast mass segmentation in comparison with FPN and LinkNet. Therefore, the traditional super-pixel-based segmentation method has shown less efficient breast mass segmentation results. This work has shown that DL methods present promising performances in comparison with traditional methods.
