Model-based segmentation and registration of multimodal medical images

Abstract

The research presented in this thesis proposed and developed a new automatic model-based registration system based on neural network techniques for CT/CT and CT/MR image segmentation/registration. Firstly, an adaptive thresholding method was proposed for CT image segmentation. With the extracted bone surface from CT images, a bone surface model was constructed using a multilayer perceptron (MLP) neural network. Secondly, a surface representation function was derived from the resultant neural network model, and then adopted for intra-operative registration. Thirdly, In CT/MR registration, the system performs CT/MR registration and MR image segmentation iteratively. The bone model was used as the reference in the proposed double-front level set MR image segmentation method. In order to reduce the possible registration error from misclassification of soft tissue surrounding the bone in MR images, a weighted surface-based registration scheme was developed. Experimental results demonstrated advantages of our method and its application to different anatomies.