Overcoming limitation of dissociation between MD and MI classifications of breast cancer histopathological images through a novel decomposed feature-based knowledge distillation method

Abstract

Magnification-independent (MI) classification is considered a promising method for detecting the histopathological images of breast cancer. However, it has too many parameters for real implementation due to dependence on input images in different magnification factors. In addition, magnification-dependent (MD) classification usually performs poorly on unseen samples, although it has lower input image sizes and fewer parameters. This paper proposes a novel method based on knowledge distillation (KD) to overcome the limitation of dissociation between MI classification and MD classification of breast cancer in histopathological images. The proposed KD method includes a pre-trained MI teacher model that is responsible for training an unprepared MD student model developed through only one magnification factor. In the proposed method, the decomposed feature maps of a teacher's intermediate layers are transferred as dark knowledge to a student. According to the experimental results, the student model developed through 40X images yielded accuracy rates of 99.41%, 99.26%, 99.14%, and 99.09% in response to unseen samples of 40X, 100X, 200X, and 400X images, respectively. Moreover, comparison results indicated the competitive performance of the proposed student model as opposed to the state-of-the-art method based on deep learning on BreakHis.

Keywords: BreakHis; Breast cancer; Convolutional neural network; Histopathology; Knowledge distillation.