Automatic Segmentation of Colorectal Polyps based on a Novel and Innovative Convolutional Neural Network Approach

Polyp is the name of a colorectal lesion which is created by cells clumping on the lining of the colon.
The colorectal polyps can lead to severe illnesses like colon cancer if they are not treated at the early stage of
their development. In current days, there are very many different polyp detection strategies based on biomedical
imageries such colon capsule endoscopy (CCE) and optical colonoscopy (OC). The CCE imagery is non-invasive
but the quality and resolution of acquired images are low. Moreover, it costs more than OC. So, today OC is the
most desired method for detecting colorectal polyps and other lesions besides of its invasiveness. To assist
physicians in detecting polyps more accurately and faster, machine learning with biomedical image processing
aspect emerges. One of the most the state-of-the-art strategies for polyp detection based on artificial intelligence
approach are deep learning (DL) convolutional neural networks (CNNs). As the categorization and grading of
polyps need significant information about their specular highlights like their exact shape, size, texture and in
general heir morphological features, therefore it is very demanded to employ semantic segmentation strategies
for detecting polyps and discriminating them from the background. According to this fact, a novel and innovative
method for polyp detection based on their semantic segmentation is proposed in this paper. The proposed
segmentation classifier is in fact a modified CNN network named as U-Net. The proposed U-Net provides an
advanced and developed semantic segmentation ability for polyp detection from OC images. For evaluating the
proposed network, accredited and well-known OC image databases with polyps annotated by professional
gastroenterologists known as CVC-ClinicDB, CVC-ColonDB and ETIS-Larib, are employed. The results of
implementation demonstrate that the proposed method can outperform the other competitive methods for polyp
detection from OC images up to an accuracy of 99% which means that the life lasting hopes could be increased
to a considerable ratio.