Xulei Yang, Si-Yong Yeo, Jia Mei Hong, Sum Thai Wong, Wai Teng Tang, Zhen Zhou Wu, Gary Lee, Sulin Chen, Vanessa Ding, Brendan Pang, Andre Choo, and Yi Su
Deep Learning, Convolutional Neural Networks, Tumor Tissue, Image Classification, Digital Pathology
The performance of automatic tumor tissue classification using traditional machine learning methods greatly depends on the choice of meaningful descriptive features derived from the tissue images. Identification of such features requires domain-specific expert knowledge and is not a straightforward process. This paper studies the feasibility of using deep learning for automatic tumor tissue classification by presenting a deep convolutional neural network (DCNN) that consists of multiple hidden layers with convolutional, max-pooling and fully connected layers. Compared to traditional machine learning methods, the proposed DCNN is able to automatically learn the layers of features to ameliorate the difficulties of prescribing features for the image classification problems. Several practical ways to tune the parameters for the DCNN are also discussed, which will be helpful for users who are keen to adopt a similar approach to their own applications. The performance of our DCNN is compared to two well-known machine learning methods – the support vector machine (SVM) and the extreme learning machine (ELM). The comparison shows that our approach achieves significant improvement in terms of classification accuracy. More importantly, the proposed approach is general and can be applied to other biomedical or biological datasets.