A Model for Object Recognition in Liver Resection Surgery

Abstract

Laparoscopic liver resection is a safer alternative to open surgery for treatment of liver cancer, a disease which claims almost 800 000 lives every year. The procedure involves making small incisions in the abdomen where instruments and a camera, called a laparoscope, are inserted. One of the major drawbacks of laparoscopic surgery is the restricted view and orientation, as well as lack of haptic feedback. Incorporating Augmented Reality, or AR, in the laparoscopic view is a proposed method of facilitating the navigation. This work extends a previous model for projecting information from 3D to 2D and vice versa using reference points, which correctly visualizes the shape, angle and size of a tumor in AR in the 2D laparoscopic view. To enable the 2D-to-3D projection, two object recognition models based on image segmentation and edge detection, respectively, were developed where white reference objects were distinguished from the darker tones of the liver tissue. The positions of the reference objects were then measured. The latter model, albeit effective given certain frames, failed to identify fiducials over the course of a test film. Since the process of image segmentation is computationally heavy, it was localized to an area of interest in a given frame, reducing the algorithm’s runtime. Statistical error estimation was used to validate the positions found by this recognition model. The average position error produced was between 1 to 5 pixels, where the frames had a pixel height of 1080. Future work involves combining the recognition algorithm with the projection model to examine the effect of the deviations of the estimated positions in the 2D laparoscopic view.