Automatic Detection Of Colonic Polyps Based On Deep Learning

A computer-aided diagnosis(CAD)system for CT colonography(CTC)can greatly improve the efficiency of doctors and greatly reduce the possibility of polyp missing.Generally,a typical CAD system is composed of two key components,polyp candidate screening and false positive(FP)reduction.In this paper a simpler solution is enabled by the application of deep learning.The three-dimensional(3D)convolutional neural network(CNN)technology was utilized to propose a simple but powerful one-step polyp detection system.With the help of the 3D fully convolutional network(FCN)technology,the step of FP reduction,the electronic colon cleansing and the accurate colon wall segmentation are no longer needed.And the detection efficiency of the proposed system was improved.A multi-step training scheme and a hybrid loss function were proposed to transform the FCN technology originally used for segmentation tasks to adapt to the specific detection task.Therefore,the pixel-level polyp segmentation annotation can be utilized to directly train the network so that information such as the size,shape and location of the polyp can be seamlessly integrated into the training.Through these improvements and the application of adaptive data transformation technology,the problem of data insufficiency for the 3D CNN application in the CAD-CTC field was alleviated to a large extent.Through experiments on public datasets,the proposed method was compared with the state-of-the-art methods in several published literatures.It was found that the method was superior in performance.Moreover,with detailed comparison experiments and quantitative evaluations,we demonstrated the effectiveness of several key modules in the proposed system.