Research On Image Segmentation And Registration In Tumor Surgery Navigation

Computer aided surgery(CAS) is image-guided intervention and one of the most discussed issues in surgical intervention field. It can guide surgeons to place lesions, make diagnosis and carry out more precise interventation through series procedure of surgical planning, surgical simulation, intra-operation registration and postoperative valiation. In this way, it tackles location problem in traditional operation and greatly reduces the risk of surgical complications. Due to the complicated structure of tissues that tumor is attached to and infiltrating character of tumor tissue, taking brain that is full of nerves and blood vessels for example, neurosurgery has imperious demands on the surgical navigation system.This thesis focuses on two major topics related to surgical navigation accuracy, which are tumor segmentation and diagnosis and registration before and in surgery. The aim of our research is to provide possibility of high accuracy surgery, relieve patients, ensure safty and save time.The main content of the thesis consists of the following major aspects:Research of automatic tumor segmentation and diagnosis algorithmAutomatic segmentation and early diagnosis of tumor provides possibility for pre-operative planning, and tackles time-comsuming and low accuracy problem of traditional manual methods. According to the above mentioned problems, this thesis proposes new methods:1. Based on traditional convolutional neural networks(CNNs), a new architecture model is proposed for automatic brain tumor segmentation and diagnosis, which combines multi-modality images. The newly designed CNNs model automaticly learns useful features from multi-modality images to combine multi-modality information. Experiment results show that the proposed model is more accurate than traditional methods and can provide reliable information for clinic treatments.2. Multi-scale CNNs algorithm is proposed based on traditional CNNs. Firstly, this algorithm utilizes self-learning mode to extract obvious features from multi-modality images, which abandons manual designed features of traditional machine learning methods. Secondly, both global and local information are combined in multi-scale CNNs overcoming infiltrating and low comparaty problem of tumor structure, while only local information is utilized in traditional methods.Research on tumor image registration methodsThis thesis proposed a new registration framework for situation that large deformation exists as deformation and continuous growing character of tumor, which induces new chanllege to registration in surgery. The proposed framework and improved pre-registration algorithm tackles corruption of traditional methods when large deformation exists. At the same time, the newly designed similarity metric accelerates convergence speed while maintaining the exact registration accuracy.1. A deep adaptive registration framework and new pre-registration algorithm based on CNNs is introduced. In traditional registration framework, pre-registration executes only once with fixed number of deformation registration. However, the proposed adaptive registration framework always executes pre-registration after every following deformation registration, greatly improving registration accuracy. As traditional pre-registration method corrups when big deformation exists between fixed and moving images, this thesis proposes to obtain rotation, translation and scale parameter separately in affine transformation of pre-registration. Firstly, CNNs classifier is trained offline to recognize rotation parameter as much as 360. As for scale parameter, image size information between fixed and moving images is utilized. Translation parameter is obtained through calculation of each image’s centroid. Experiment result shows that method in this thesis can solve problem when large deformation exists instead of traditional registration framework.2. An effective similarity metric is introduced. Principle component analysis(PCA) is utilized to extract the main feature points of both fixed and moving imgaes in registration, avoiding accuracy influences caused by noises. Through combination of PCA and traditional similarity metrics such as Spearman and Pearson, etc., new similarity metric forms. Experiment result shows that this algorithm can greatly accelerate convergence speed with maintaince of registration accuracy.