A Comparison of Stereo-Based Methods of 3D Reconstruction for Surface-Based Registration

This work aims to study the different stereo-vision based methods of 3D reconstruction for surface registration which is used in Image guided surgeries. The study evaluates the accuracy of four different classical methods, the best of which is compared to the deep learning based LEAStereo model for the surface reconstruction. For this study a phantom of an arbitrary patient's head was taken with known ground truth geometry. The reconstructed surfaces are compared with the CT scan of the head phantom to evaluate the accuracy of the reconstruction. The evaluation results provide insights into the accuracy, density, computation time and performance of all the four classical methods of reconstruction. Then the best among the classical methods is then exclusively compared with the deep learning method LEAStereo. Overall, this study contributes to the understanding of the 3D reconstruction of the surface using stereo-vision based methods and the feasibility for surface registration in medical imaging.