Automatic Polyp Detection in Colonoscopy Images: Convolutional Neural Network, Dataset and Transfer Learning

摘要

Colonoscopy plays an essential role in colorectal cancer prevention and diagnosis. Due to the high miss-rate of colon polyps the application of automated polyp detection technology in clinical is necessary. However, despite researchers made significant progress, automatic polyp detection is still a challenge task. In recent years, deep learning shines in medical image processing and achieved satisfactory result in different kinds of medical images. In this paper, we propose an end-to-end convolutional neural network (CNN) framework to deal with this challenge problem. The whole framework consists of 16 convolutional layers and 6 pooling layers. In order to improve the performance of the proposed method we employ transfer learning algorithm to fine tune the pre-trained model. Several effective tricks in deep learning also adopt to train the network we proposed. Compared with other methods employing traditional algorithms or hand-crafted features, CNN has the ability to reach lower error rate and faster speed. More importantly, we establish a novel colonoscopy dataset to train our neural network. The dataset consists of more than 10 thousand high resolution images which are carefully selected from hospital. We evaluate our classification system using precision, recall and F1 score analysis. The final model obtained 95.2% precision, 97.9% recall and 96.5% F1 score. In addition, we draw receiver operation characteristic (ROC) curve and the area under ROC curve can reach 96.6%. For location task, we employed Intersection over Union (IoU) to evaluate the model and get 0.65 IoU score.