Deep quaternion convolutional neural networks for breast Cancer classification

Breast Cancer nowadays has been a major cause of death in women worldwide and this has also been confirmed by the World Health Organization. The severity of this disease can be minimized to a large extent if it is diagnosed properly at an early stage. Two important types of tumors found in the case of breast cancer are malignant and benign. Moreover, It has been observed that, unlike benign tumors, malignant tumors are more dangerous because of their invasive nature. Therefore, the proper treatment of a patient having cancer can be processed in a better way, if the type of tumor can be identified as early as possible. Deep neural networks have delivered a remarkable performance for detecting malignant tumors in histopathological images of breast tissues. However, the existing works today, are focused much on real-valued numbers. When data is multi-channel such as images and audio, conventional real-valued CNN on flattening and concatenating loses spatial relation within a channel. To address the above-said issues, we have exploited a quaternion residual network for detecting breast cancer in a dataset of histopathological images, which are publically available in the dataset of Kaggle. In this work, we first transform breast histopathological images into quaternion domains. Second, the Residual CNN was customized to work in the quaternion domain so that it extracts the better representative features for multidimensional input objects. Extensive experimental results demonstrate that our model architecture although takes slightly more time to train but it offers an increased classification accuracy of 97.20% which is more than the performance of a residual network compatible with real numbers. Also, the proposed model outperforms when compared against the baseline neural network models.