Histological breast cancer classification using CNN and MLP based ensembles

Abstract

Breast cancer affects thousands of people worldwide each year, artificial intelligence used for digital pathological computer-aided diagnosis for breast cancer classification is a valuable domain. The advancement of machine learning and deep learning methods, as well as pathology slide digitization for primary diagnosis, is a major development toward meeting the demand for more accurate breast tumor diagnosis, classification, and prediction. This paper proposes and evaluates a new approach consisting of deep hybrid homogenous ensemble method based on seven deep learning models for feature extraction (DenseNet 201, Inception V3, Inception ResNet V2, MobileNet V2, ResNet 50, VGG16, and VGG 19), a multi-layer perceptron (MLP) for classification and two combination rules (hard and weighted voting) for histological classification using the BreakHis dataset four magnification factors: 40X, 100X, 200X and 400X. The outcomes proved the potential of the proposed new approach since it outperformed its singles achieving an accuracy value of 98,3% for the MFs 40X and 100X and 97,7% for the MFs 200X and 400X.