Graph-search Based UNet-d For The Analysis Of Endoscopic Images

Abstract

While object recognition in deep neural networks (DNN)
\nhas shown remarkable success in natural images, endoscopic
\nimages still cannot be fully analysed using DNNs, since
\nanalysing endoscopic images must account for occlusion,
\nlight reflection and image blur. UNet based deep convolutional
\nneural networks (DNNs) offer great potential to extract
\nhigh-level spatial features, thanks to its hierarchical nature
\nwith multiple levels of abstraction, which is especially useful
\nfor working with multimodal endoscopic images with white
\nlight and fluoroscopy in the diagnosis of esophageal disease.
\nHowever, the currently reported inference time for DNNs is
\nabove 200ms, which is unsuitable to integrate into robotic
\ncontrol loops. This work addresses real-time object detection
\nand semantic segmentation in endoscopic devices. We
\nshow that endoscopic assistive diagnosis can approach satisfy
\ndetection rates with a fast inference time.