Image Segmentation Methods in Mammography: a Literature Review

Resumen

This article aims to present a systematic literature review on modern image segmentation methods in mammography. Since the 1980s, mammography has been a standard examination used to identify nodules and signs of breast cancer, being an accessible and low-cost examination. Currently, breast cancer is the most frequent type of cancer in women in Brazil. As a form of early disease control, imaging exams are an important medical ally, and in recent years technology has been used to identify lesions in images. This work was restricted to research articles that use image segmentation methods involving machine learning and deep learning. The selection adopted was of works published in the last five years in the journals: Elsevier, IEEE, Springer and MDPI. Initially, 550 articles were selected, and of these, only 51 passed the funneling stage. The literature review revealed a predominant use of convolutional neural networks (CNNs), mainly U-Net and ResNet, and free databases, especially CBIS-DDSM and DDSM. In recent years, significant results have been observed with the U-Net architecture and with CNN+ Transformers integrative models. The accuracy of the models exceeds 98% in some studies, and recently, the FDA approved the use of deep learning technologies to support medical decision-making.