Liver segmentation in color images

We describe the use of a deep learning method for semantic segmentation of the liver from color images. Our intent is to eventually embed a semantic segmentation method into a stereo-vision based navigation system for open liver surgery. Semantic segmentation of the stereo images will allow us to reconstruct a point cloud containing the liver surfaces and excluding all other non-liver structures. We trained a deep learning algorithm using 136 images and 272 augmented images computed by rotating the original images. We tested the trained algorithm on 27 images that were not used for training purposes. The method achieves an 88% median pixel labeling accuracy over the test images.