Non rigid registration of 3D MR images

Abstract

With the advent of interventional radiology and modern radiation treatment, the radiological images have found widespread usage in clinical trials and medical research. Therefore the information gained from different images must be properly integrated to aid in diagnosis and treatment. The analysis procedure starts with the geometric alignment or registration of several images which are to be used in treatment and surgical planning. The registration process gets simple with rigid body assumption, but techniques employing this approach have limited applicability. Moreover, it cannot model the organ deformations and other changes in anatomy and pathology. Non-rigid registration is a key enabling technique that can model these locally variant deformations. A non-rigid registration algorithm is proposed here that is automatic, accurate and computationally efficient. This non-rigid registration method is then used to segment the subcortical structures in subject MR image. The results of automatic and manual segmentation were contrasted using positive predictive value, sensitivity and Dice coefficient metrics.