Evaluation of Methods for Semantic Segmentation of Endoscopic Images

We examined multiple semantic segmentation methods, which consider the information contained in endoscopic images at different levels of abstraction in order to predict semantic segmentation masks. These segmentations can be used to obtain position information of surgical instruments in endoscopic images, which is the foundation for many computer assisted systems, such as automatic instrument tracking systems. The methods in this paper were examined and compared in regard to their accuracy, effort to create the data set, and inference time. Of all the investigated approaches, the LinkNet34 encoder-decoder network scored best, achieving an Intersection over Union score of 0.838 with an inference time of 30.25 ms on a 640 x 480 pixel input image with a NVIDIA GTX 1070Ti GPU.