Deep Quaternion Residual Learning for Breast Cancer Classification

Abstract

Convolution neural networks (CNN) have shown the state of the art performance for visual recognition, classification of images, time series, and sequential data. Despite good performance, it has few serious drawbacks that it cannot properly encode the orientation and spatial positioning of components of the input data and it also suffers from overfitting. Quaternion CNN is a generalization of traditional CNN and it can properly encode internal and external dependencies between the components of the input data and it is free from overfitting and its generalization performance is better than conventional CNN. It can be modified to work with all Deep Neural Network (DNN) models with quaternion as input. In this paper, the authors have proposed quaternion residual learning for the classification of breast cancer in the BreakHis dataset of breast histopathological images. They have observed that quaternion CNN outperforms to real CNN. The experiment has obtained the classification accuracy of 98.04% and F-score of 0.9842.