Deep Stacked Ensemble for Breast Cancer Diagnosis

Breast cancer is considered one of the major public health issues and a leading cause of death among women in the world. Its early diagnosis can significantly help to increase the chances of survival rate. Therefore, this study proposes a deep stacking ensemble technique for binary classification of breast histopathological images over the BreakHis dataset. Initially, to form the base learners of the deep stacking ensemble, we trained seven deep learning (DL) techniques based on pre-trained VGG16, VGG19, ResNet50, Inception_V3, Inception_ResNet_V2, Xception, and MobileNet with a 5-fold cross-validation method. Then, a metamodel was built, a logistic regression algorithm that learns how to best combine the predictions of the base learners. Furthermore, to evaluate and compare the performance of the proposed technique, we used: (1) four classification performance criteria (accuracy, precision, recall, and Fl-score), and (2) Scott Knott (SK) statistical test to cluster and identify the outperforming models. Results showed the potential of the stacked deep learning techniques to classify breast cancer images into malignant or benign tumor. The proposed deep stacking ensemble reports an overall accuracy of 93.8%, 93.0%, 93.3%, and 91.8% over the four magnification factors (MF) values of the BreakHis dataset: 40X, 100X, 200X and 400X, respectively.