Artifact Detection in Endoscopic Video with Deep Convolutional Neural Networks

Gastrointestinal cancer is a common and deadly disease that affects many people in the world. In 2019, Gastrointestinal cancer was the most common cancer and the second leading cause of death in the US. Detecting gastrointestinal cancer during the early stage is the most effective way to improve the survival rate. One of the commonly used clinical procedures for early detection of gastrointestinal cancer is endoscopy. The main challenge of a high-quality endoscopy operation is the presence of various forms of artifacts during the operation, e.g., pixel saturation, motion blur, defocus, specular reflections, bubbles, fluid, debris. These artifacts not only increase the difficulty in examining the underlying tissues during diagnosis but also affect the post-analysis methods required for follow-ups (e.g., video mosaicking for follow-ups and archival purposes and video-frame retrieval for reporting). Also, the presence of these artifacts often interferes with the computer-aided diagnosis of various lesions in endoscopy. The Convolutional Neural Network (CNN) based object detection methods have proved to be an effective approach for nature image object detection and colonoscopy applications (e.g., polyp detection). However, fewer efforts have been devoted to endoscopic artifact detection due to the lack of training data. In this paper, we use data from the EAD2019 challenge and investigate the performance of two improved CNN-based methods for seven-class endoscopic artifact detection (EAD). Experiment results show that our proposed objection detectors based on SSD and Faster-RCNN significantly outperform the baseline.