Investigation and improvement of multi-resolution mutual-information-based registration
Handini Dilla · 2006
An imageguided surgical robotic system, NeuroBot, is currently being developed by a team of researchers from Nanyang Technological University (NTU) Singapore and National Neuroscience Institute (NNI) Singapore.The system aimed to assist neurosurgeons in skullbase surgery especially in generating operative plan and drilling task.As one of the required functionalities of the system, an image registration technique was aimed to be developed in this project.MutualInformation (MI)based method, an acknowledged registration method for its high accuracy, was adopted in a multiresolution scheme to improve the speed and robustness.It was intended to use multimodal images in this project.However, due to the lack of availability of public resources in multimodal datasets, the project used single modal image datasets from MRI (Magnetic Resonance Imaging).Two versions of algorithms adopting the technique were developed, and their performance were investigated for several factors which are optimization methods, number of multi resolution levels, transformation, and acquisition modes.The speed and accuracy served as the performance indicator.The algorithms differ by their optimization methods which are Powell's and regular step gradient descent (RSGD) optimizations.ITK (Insight Toolkit) registration framework was used as the development tool kit.It was found that RSGD outperforms Powell's method -based algorithm in terms of speed, with comparable accuracy.The experimental results using RSGD optimization registration technique showed that only two results, out of 94 results, exceeded the subvoxel accuracy; such results are termed as outlier data (in this case, data whose error > 1 mm).Multiresolution scheme was proven to improve the robustness and speed of registration process; and, it was found that the number of multiresolution levels affects the registration speed.In the experiments, among the number of levels tested, three and four levels were found to be the optimum number of levels as they gave the fastest speed for RSGDbased algorithm.It was found that time spent for image resampling for every level acts as the limiting factor to the registration speed.The placement of the transformation origin was also found to affect the registration speed.The placement of transformation origin in the center (centeredorigin) of an image volume gave faster registration process than the one in the image corner (normalorigin) by four times.In addition, the experiment showed