Transfer and ensemble learning models in breast mammogram pathology classification

Abstract

Breast cancer is a global concern disease that specially affects women. Early detection of the disease,

through mammography, increases life expectancy and reduces serious consequences of it. However, the

sensitivity of mammography is variable; especially in dense breast. Therefore, it is important to develop

cost effective tools that help to reduce the false positive and negative rates in the radiologist's diagnosis by

providing a second opinion. Since year 2012, Deep Learning has achieved optimistic results in image

classification, object detection, and image segmentation through convolutional neural networks. This

master's thesis explores transfer learning, fine tuning, whole retrain, and ensemble learning techniques to

classify mammogram masses as benign or malignant. For this purpose, 20 pre-trained convolutional neural

networks on the ImageNet Dataset were compared. Furthermore, this work's proposed Automatic Soft

Voting Ensemble (ASVE), which used a Perceptron to automatically tune the soft voting weights of the

fine tuned convolutional neural networks ensemble, has achieved promising results on Inbreast (AUC =

98.2%), MIAS (AUC=97.8%), and in a merged dataset of both (AUC=91.8%). This suggests that this

work's approach is suitable to be implemented in a computer aided diagnostic tool.