Detecção de Metástase em Câncer de Mama usando Ensemble Deep Learning

Abstract

Breast cancer is a disordered growth of breast cells, which unnormal evolution can result in metastasis, the worst stage of this issue. To avoid cancer growth, a fast diagnostic is required. Thus, researchers worldwide are investigating and developing technological solutions to detect breast cancer with higher precision, compared with the results of pathologists, which can reach 0,72 AUC to analyze exams with thousands of images at diferent angles. The state-of-art works integrate concepts of Deep Learning with the application of Convolutional Neural Networks with other algorithms or architectures combined, including Machine Learning techniques, leading to promising results equal to or more than 0.90 AUC. This work presents an algorithm that applies Ensemble Convolutional Neural Networks and Transfer Learning using data from the PatchCamelyon database (composed of metastatic histopathologic images of lymph node sections of breast cancer). Results show that the proposal reaches 0,9565 of AUC and a loss rate of 0,2869 with the combination of U-Net and VGG19 architecture