Multiple Magnification Learning: A Deep Learning Approach For Classifying Breast Tumor From Histopathological Images Based On The Concept Of Multiple Instance Learning

Abstract

The classification of breast cancer tumors in histopathological images is an expanding research field in computer-aided diagnosis. While the conventional approach to breast tumor classification involves extensive labor in hand-engineering features used for classification, the development of deep learning techniques has enabled a new approach that requires less human involvement in feature extraction. The Convolutional Neural Network (CNN) is one of the deep learning techniques that allows for the learning of feature representation given only input images. The development of CNNs has been common in many computer vision problems and is recently employed in histopathological image analysis. This research proposes an approach of Multiple Magnification Learning (MML) that utilizes the concept of Multiple Instance Learning (MIL). It employs a deep Convolutional Neural Network (CNN) model with four input paths, which simultaneously processes images at four different magnification levels. The backbone network used in the model is EfficientNetV2-S for histopathological image classification. The proposed method significantly outperforms previous state-ofthe-art approaches in terms of 97.12% accuracy, 97.21% precision, 98.59% recall, 97.89% F1-score, 93.35% Matthew’s Correlation Coefficient (MCC), and Area Under the Curve (AUC) of 99.78% when evaluated on an independent test dataset. A web application with user-friendly interface is also designed to assist the histopathologists.