A combined Feature extraction technique for cancer classification based on deep learning approach

Extraction from large no. of features (genes signatures) are the major issues in the prediction of cancer and its specific type identification using microarray datasets. Even though most classifiers predict the class (normal or cancerous) for various cancers, the accuracy of prediction still suffers. This is due to the importance of fewer gene signatures for a particular cancer and classification of samples independent from their originating form. The present paper proposes gene extraction techniques to work in an unsupervised manner. The proposed technique takes the advantage of both linear and non-linear feature extraction methods. Principal component analysis (PCA) is used in a linear manner whereas Denoising Autoencoder (DAE) is used in a nonlinear manner. In the first phase of the work, feature space extracts from both the methods have been combined and new features space has been utilized for cancer classification. Here, Four classifiers: Support vector machine (SVM), Multilayer perceptron (MLP), Naive Bays (NB) and Decision Tree are applied on a no. of gene signatures extracted from four different cancer datasets. It is seen that after feature extraction from this PCA-DAE, classification accuracy either increases or attains its maximum value in comparison with base techniques except NB.