Deep and dense convolutional neural network for multi category classification of magnification specific and magnification independent breast cancer histopathological images

Abstract

Nowadays, breast cancer is the leading cause of mortality among all other cancers affecting women. Histopathological images serve as an appropriate template for the pathological investigation of this disease. However, a proper and precise classification of these images is indispensable for an accurate prognosis. Over the time numerous attempts have been made for classification of histopathological images. However, most of the research works produced on this subject deal with identifying benign and malignant categories only. Since specific treatments and therapeutic schedules exist for different sub-categories of this disease, identifying the actual sub-class on a sample image is highly desirable. The present study proposes a deep learning based framework for automatic classification of breast cancer images into sub-classes of benign and malignant categories performed on magnification dependent and magnification independent samples of the BreakHis dataset. The proposed framework based on DenseNet architecture achieved respective mean patient recognition scores of 96.55% and 91.82% for binary and multi-classifications of H&E stained images of the BreakHis dataset. Likewise, at the image level mean recognition scores of 91.72% and 96.72% achieved by the proposed framework outperform the results reported in various latest available studies employing deep convolutional neural networks and state-of-the-art architectures. Further, the generalization ability of the proposed framework is tested on a different histopathological dataset - ICIAR 2018, yielding competitive accuracy scores of 93.25% and 92.3% at patient and image levels respectively.