Abstract

The massive quantity of data delivered by whole slide images for breast cancer diagnosis necessities building highly efficient deep learning pipelines to automate patches classification. Through ensemble learning, histology image classification aims to detect breast cancer areas automatically in breast biopsy specimens. The present studies concentrated on conventional ensemble methods, such as majority voting, that over-simplify the problem. Traditional ensemble methods presume uniformity of weights of deep features, often leading to inaccurate predictions. This paper proposes a Neural Network ensemble that learns the weights of different deep features to derive an image-wise prediction. The deep learning features are acquired via fine-tuning pre-trained models. We experimented with our approach on two publicly available breast histology images to verify the effective- ness of our method. Our proposed methods significantly outperformed recently published results on both datasets in terms of accuracy and other state-of-the-art evaluation metrics.
Full Paper