A combined feature-vector based multiple instance learning convolutional neural network in breast cancer classification from histopathological images

Breast cancer, being one of the leading causes of death in women, has a dangerous impact worldwide. Thousands of lives can be saved if the detection of breast cancer at an early stage is possible. Pathologists, after analyzing the morphological characteristics of biopsy samples at different levels of magnification, identify the existence of cancerous tissue. This total diagnostic process depends on the subjective analysis of pathologists and it may lead to some difficulties. In this situation, Computer-Aided Diagnosis (CAD) might be quite useful. The goal of this research is to develop a CAD system to be used as an assistant to the pathologists in making the final decision to diagnose breast cancer accurately. In this paper, our study is conducted on a publicly available dataset called BreakHis in which the histopathological images are categorized into four magnification levels. The concept of Multiple Instance Learning (MIL) is applied in this research. A deep Convolutional Neural Network (CNN) model, with four input paths, is used to take the images at four different magnification levels parallelly. EfficientNet-B0 is applied as the backbone network in our model to classify the histopathological images. Our proposed approach surpassed previous state-of-the-art works by a significant margin in terms of accuracy, precision, recall, F-measure, Matthew's Correlation Coefficient (MCC), and Area Under the Curve (AUC) when applied to an independent test set.