Fully Convolutional DenseNets for Polyp Segmentation in Colonoscopy

Abstract

Early diagnosis and resection of colorectal polyps can effectively reduce the incidence and mortality rate. Colorectal cancer is a common gastrointestinal malignancy, ranking one of the three major malignancies around the world. With the improvement of living standards and dietary habits related problems, the incidence and mortality of colorectal cancer are showing an upward trend. Colorectal cancer is mostly from adenoma polyp malignant change, so early detection has important clinical significance. Although colonoscopy conducted by doctors is considered the most effective way in detecting polyps, uncertainty such as fatigue can affect the results. To solve this problem, we propose a fully convolutional densenet method to achieve the automatic detection and segmentation of colorectal polyps by computer. In this paper, we apply densenet to full convolutional network in segmentation of colorectal polyp, under the condition that not requiring post-processing and pre-training situation, we compare the number of parameters in different layers and assess accuracy and IOU respectively in segmentation of colorectal polyps. The results show that accuracy is improved as the layer increases gradually. When the layer number is 78(N=78), accuracy reaches 97.1% and the average IOU is 83.4%. In addition, we make a comparison with the state-of-the-art polyp segmentation method, the results reveal our method make a considerable improvement.