3D Feature Extraction Based On Sift And Its Application In Medical Image Registration

Image registration is a process of establishing spatial correspondences betweentwo images. It is widely used in computer vision, remote sensing data analysis andimage processing. Especially in image-guided radiation therapy, image registrationplays an important role. Accurate and effective registration methods can helpdoctors carry out medical diagnosis, make operation planning, trace tumor position,delineate therapeutic area and evaluate effectiveness of the treatment.In general, image registration can be divided into intensity-based andfeature-based methods. However, intensity-based methods face two kinds ofpractical problems: high computational cost and misregistration of fine tissuestructures. Features are very important characteristics of any medical images andcan be employed in nonrigid image registration. SIFT is considered as one of themost effective feature extraction methods and has been used in medical imageregistration, and promising results has been obtained by using it. However, SIFT isapt to detect blob features which cannot reflect properly the motion of tissues. Inthis work, we proposed a hybrid feature detection method which can detect tissuefeatures effectively based on Harris and SIFT, and the effectiveness was proved byexperiments.The SIFT feature extraction algorithm is studied in depth in this dissertation. Inorder to extract tissue features effectively, we proposed a hybrid feature detectionmethod based on Harris and SIFT. The feature points detected by this methodmostly distribute in the vessel intersections and organizational boundaries, whichcan reflect the deformation accurately.However, there are some wrong matched pairs after feature matching process.To remove these mismatched points, we proposed a novel method based oncross-correlation and structural invariance. Then, a TPS transformation is employedto establish the registered image using corresponding point pairs.A series of thoracic CT images are tested by using the proposed algorithm, andthe quantitative and qualitative evaluations show that our method is much betterthan conventional SIFT method, especially in the case of large deformation of lungsduring respiration.