Localization of Polyps in WCE Images Using Deep Learning Segmentation Methods: A Comparative Study

Abstract

Wireless capsule endoscopy (WCE) is a promising technology for the investigation of gastrointestinal (GI) tracts. Long WCE video pose a challenge before the field experts and therefore automatic identification of anomalous video frames and localization of anomalies is an important research problem. Various segmentation models based on deep learning techniques for object localization have been suggested for different type of applications including biomedical applications. Recently, some researchers have applied these methods for the analysis of WCE images. In this paper four popular deep learning-based segmentation models UNet, SegNet, PSPNet, and Fully Convolutional Network (FCN) are analyzed for their suitability in localization of anomaly. Popular CNN architectures like VGG16, MobileNet, and ResNet50 are used as the base architecture to create the segmentation models. In this work two well known polyp datasets namely ETIS-Larib and CVC-ClinicDB are considered. The outcomes show that UNet and SegNet developed using MobileNetV1 architecture are performing relatively better than other architectures. Mean IoU of 0.910 and 0.883 is achieved using MobileNet-UNet whereas using MobileNet-SegNet it reaches 0.894 and 0.885 on CVC-ClinicDB and ETIS-Larib datasets respectively.