Clinical outcome of several Minimally Invasive Surgeries (MIS) heavily depend on the accuracy of intraoperative pose estimation of the surgical instrument from intraoperative X-rays. The estimation consists of finding the tool in a given set of X-rays and extracting the necessary data to recreate the tool's pose for further navigation - resulting in severe consequences of incorrect estimation. Though state-of-the-art MIS literature has exploited image registration as a tool for instrument pose estimation, lack of practical considerations in previous study design render their conclusion ineffective from a clinical standpoint. One major issue of such a study is the lack of Ground Truth in clinical data -as there are no direct ways of measuring the ground truth pose and indirect estimation accumulates error. A systematic way to overcome this problem is to generate Digitally Reconstructed Radiographs (DRR), however, such procedure generates data which are free from measuring errors (e.g. noise, number of projections), resulting claims of registration performance inconclusive. Generalization of registration performance across different instruments with different Degrees of Freedom (DoF) has not been studied as well. By marrying a rigorous study design involving several clinical scenarios with, for example, several optimizers, metrics and others parameters for image registration, this paper bridges this gap effectively. Although the pose estimation error scales inversely with instrument size, we show image registration generalizes well for different instruments and DoF. In particular, it is shown that increasing the number of X-ray projections can reduce the pose estimation error significantly across instruments - which might lead to the acquisition of several X-rays for pose estimation in a clinical workflow.
展开▼
