Deep Oral Cancer Lesion Segmentation with Heterogeneous Features

Abstract

About 650,000 new cases of oral cavity cancer occur every year in the world, and cause more than 330,000 deaths. If oral cancer is diagnosed at an early stage, the overall 5-year survival rate is over 70%, while it drops to less than 40% if detected at a late stage. Thus, early detection of oral cancer is important. Visual non-invasive examination is an efficient and feasible approach for performing a preliminary diagnosis of oral cancer. In this paper, we propose a fully convolutional network (FCN) based model to segment cancer and precancer lesion regions in the oral cavity. In addition to the RGB channels of the input image, we append features of Gabor filter and wavelet filter that show strong response at cancer and precancer regions. We also propose a refine stage before the decision layer of FCN to preventing weight dominating problem when reducing high dimension features to small number of classes. In the experiments on oral cancer dataset, the IOU, sensitivity, and specificity of the proposed network achieves 0.586, 0.883, 0.726 respectively. The experimental results show the effectiveness of our method.