Hybrid model for 2-D rigid multimodal registration of brain images

In this paper, a robust control points based 2-D rigid multimodal registration technique for brain images is proposed using weighted feature set approach (WFSA). The proposed method uses both area and feature based techniques for control points mapping. Firstly, highly descriptive rigid control points are detected in reference image using area-based technique. This reduces the search space for feature based mapping. Five different descriptive features are extracted from images and WFSA approach is applied. The WFSA assigns more weight to the more distinctive features between two images providing an efficient mapping criterion for multi-modal image registration. This enhances the computational efficiency, provides optimal location of the final control points and hence minimizes the overall localisation error. The method has been tested on two real multi-modal medical image datasets. First dataset is the standardised registered dataset from MGH (Massachusetts General Hospital) consisting of 18 registered pairs of T1 and T2 images. Geometric Distortions like scaling, translation, rotation and Gaussian random noise are randomly added in the image to be registered (target image) for validation of the proposed method. Normalized Landmark Location Error(NLLE) is used as an evaluation criterion for checking the robustness and efficiency of the method. It is observed from the experimentation that even at high degree of subject movements (-50° and +50°) and at high Gaussian random noise, NLLE less than 1% is obtained. Second dataset is collected from Post Graduate Institute of Medical Education and Research(PGIMER), Chandigarh, India.This dataset consists of unregistered pairs of MR-T1, T2, and post contrast T1 and CT imagesof 12 patients diagnosed with brain tumor. Clinically relevant information obtained after registration i.e. radiological and pathological information obtained from the registered datasets is discussed. The proposed algorithm registers the images in 55 seconds irrespective of the modality.