Breast Cancer Histopathology Image Classification Using Frequency Attention Convolution Network.

The existing deep learning works mainly capture breast cancer histopathology image features in the spatial domain, and they rarely consider the frequency domain feature representation of histopathology images. According to the classical digital signal processing theory, frequency domain features may outperform spatial domain features in analyzing texture images. Motivated by this, we attempt to mine frequency domain features for the breast cancer histopathology image classification application, and further propose a novel frequency-attention convolutional network called SeFFT-Net by combining the Fourier transform with the channel attention mechanism. The core of SeFFT-Net consists of a newly constructed frequency-based squeeze and excitation (SeFFT) module, which first performs Fourier transform with residual construction to capture deep features in the frequency domain of histopathology images, followed by a squeeze-and-excitation attention operator to further enhance important frequency features. We extensively evaluate the proposed SeFFT-Net model on the public BreakHis breast cancer histopathology dataset, and it achieves the optimal image-level and patient-level classification accuracy of 98.67% and 98.16%, respectively. Meanwhile, ablation studies also well demonstrate the effectiveness of introducing frequency transforms for this medical image application.