Deep-Hist: Breast cancer diagnosis through histopathological images using convolution neural network

Abstract

Breast cancer diagnosis utilizes histopathological images to get best results as per standards. For detailed diagnosis of breast cancer, microscopic analysis is necessary. During analysis, pathologists examine breast cancer tissues under different magnification levels and it takes a long time, can be hampered by human interpretation and requires expertise of different magnifications. A single patient usually requires dozens of such images during examination. Since, labelling the data is a computationally expensive task, it is assumed that the images for all patients have the same label in conventional image-based classification and is not usually tested practically. In this study, we are intending to investigate the significance of machine learning techniques in computer aided diagnostic systems based on analysis of histopathological breast cancer images. Publicly available BreakHis data set containing around 8,000 histopathological images of breast tumours is used for conducting experiments. The recently proposed non-parametric approach is proven to show interesting results when compared in detail with machine learning approaches. Our proposed model ’Deep-Hist’ is magnification independent and achieves > 92.46% accuracy with Stochastic Gradient Descent (SGD) which is better than the pretrained models for image classification. Hence, our approach can be used in processing data for use in research and clinical environments to provide second opinions very close to the experts’ intuition.