Wavelet Scattering Networks of Deep-Learning Features for Robust Discovery of Protein Markers in Rectal Cancer

Abstract

Cancer biomarkers play a very important role for the disease diagnosis, prognosis, and prediction of treatment response. In this study, expression levels of RhoB and DNp73 proteins in rectal cancer captured on immunohistochemical images were analyzed for predicting the 5-year survival time of two patient cohorts with and without preoperative radiotherapy. Applications of deep convolutional neural networks in medicine have been increasingly attracting attention in clinical cancer research. The success is mainly due to the power of extracting deep features of images that are very useful for machine learning. Another successful convolution-based method for feature extraction at multiple levels is wavelet scattering networks. This study combines the potential of these two convolution-based methods for robust image feature extraction of protein expression. The proposed approach was tested with samples of various tissue types (tumor, biopsy, metastasis, and normal condition adjacent to tumor). Statistical performance measures show very high rates in terms of prediction, classification and model accuracies, precision, and area under the receiver operating characteristic curve. The results suggest the usefulness of the double convolutional learning for assisting clinical researchers in timely validating and discovering cancer biomarkers.