Surface Reconstruction from Tracked Endoscopic Video Using the Structure from Motion Approach

The lack of 3D vision and proper depth perception associated with traditional endoscopy significantly limits the quality of the diagnostic examinations and therapy delivery. To address this challenge, we propose a technique to reconstruct a 3D model of the visualized scene from a sequence of spatially-encoded endoscopic video frames. The method is based on the structure from motion approach adopted from computer vision, and uses both the intrinsic camera parameters, as well as the tracking transforms associated with each acquired video frame to calculate the global coordinates of the features in the video, and generate a true size 3D model of the imaged scene. We conducted a series of phantom experiments to evaluate the robustness of the proposed method and the accuracy of a generated 3D scene, which yielded 1.7±0.9 mm reconstruction error. We also demonstrated the application of the proposed method using patient-specific endoscopic video image samples acquired during an in vivo gastroscopy procedure.