Fully Automatic Polyp Detection Based on a Novel U-Net Architecture and Morphological Post-Process

Abstract

Colorectal lesions known as polyps are one of the diagnostic symptoms for colorectal disease. So, their accurate detection and localization based on a computer-aided diagnosis can assist colonists for prescribing more effective treatments. The computer vision and machine learning methods like pattern recognition and deep learning neural networks are the most popular strategies for automatic polyp detection purpose. The proposed approach in this paper is an innovative deep learning neural network. The proposed network has a novel U-Net architecture. The architecture of proposed network includes fully 3D layers which enable the network to be fed with multi or hyperspectral images or even video streams. Moreover, there is a dice prediction output layer. This type of output layer employs probabilistic approaches and benefits from more accurate prediction abilities. The proposed method is applied to international standard optical colonoscopy datasets known as CVC-ClinicDB, CVC-ColonDB and ETIS-Larib. The implementation and evaluation results demonstrate that the proposed U-Net outperforms other competitive methods for automatic polyp detection based on accuracy, precision, recall and F-Score criteria. The proposed method can assist experts and physicians to localize colonial polyps with more accuracy and speed. In addition, the proposed network can be used on live colonoscopy observations due to its high performance and fast operability.