Target segmentation and tracking of medical endoscopic images

Abstract

Endoscopic imaging is an important tool for living detection, and it provides first-hand information for pathological diagnosis, such as digestive tract and respiratory tract disease. However, the imaging quality of endoscopic images faces many challenges due to the particularity of human internal environment. For example, water vapor and blood stains will block the lens, and unstable movements of the endoscope equipment will cause serious jitter of the field of view. In addition, the imaging range of the endoscope is limited, and the jitter often caused the problem of target loss. Based on the above problems, the algorithm on target segmentation and tracking of medical endoscopic images is proposed to assist doctors in real-time analysis, and extract accurate information in the harsh observation environment. Target segmentation and tracking are two important tasks in computer vision. Most tracking algorithms use target detection to ensure real-time performance, however, the combination of segmentation and tracking is conducive to improving the identification accuracy. In addition, most segmentation and tracking focus on the real scenarios, the related datasets include DAVIS (Densely-Annotated Video Segmentation), OTB (Online Objective Tracking Benchmark) and VOT (Visual Object Tracking). However, the endoscope images are quite different due to the small detection range and difficult control of the viewing angle. Inspire by SAT (State-Aware Tracker), AlexNet is used to perform the tracking task to ensure real-time tracking, and excessive receptive field and random sampling are introduced to adapt to the characteristic of fast speed and irregular direction of target movements in the endoscope scene. In addition, this paper uses the training enhancement of the first-frame deformation to simulate the process of target appearance. VOT and DAVIS are used for pre-training, and the Hyper Kvasir dataset, which is the dataset on endoscopic polyp, is used for semi-supervised training and performance evaluation. The Dice of the proposed algorithm is better than the comparison algorithm. The algorithm on target segmentation and tracking of medical endoscopic images can monitor the images in the entire process, and give doctors reminders and information confirmation. The fully automatic computer vision algorithm can make full use of every detail in the endoscopic surgery, and it can free doctors form the high-intensity image monitoring, so that they can put their time and energy into more important work.