Automated Polyp Segmentation in Colonoscopy Frames Using Fully Convolutional Neural Network and Textons

In this paper, we presented a novel hybrid classification based method for fully automated polyp segmentation in colonoscopy video frames. It contains two main steps: initial region proposals generation and regions refinement. Both machine learned features and hand crafted features are taken into account for polyp segmentation. More specifically, the hierarchical features of polyps are learned by fully convolutional neural network (FCN), while the context information related to the polyp boundaries is modeled by texton patch representation. The FCN provides pixel-wise prediction and initial polyp region candidates. Those candidates are further refined by patch-wise classification using texton based spatial features and a random forest classifier. The segmentation results are evaluated on a publicly available CVC-ColonDB database. On average, our method achieves 97.54% of accuracy, 75.66% of sensitivity, 98.81% of specificity and DICE of 0.70%. The fast execution time (0.16 s/frame) demonstrates the promise of our method to be used in real-time clinical colonoscopic examination.