Clustered Redundant Keypoint Elimination SURF Method in MRI Image Registration Based on Alpha-Trimmed Relationship

Zahra  Hossein-Nejad; Mehdi  Nasri

doi:10.18502/fbt.v10i2.12216

Zahra Hossein-Nejad Department of Electrical Engineering, Sirjan Branch, Islamic Azad University, Sirjan, Iran
Mehdi Nasri Department of Electrical Engineering, Khomeinishahr Branch, Islamic Azad University, Isfahan, Iran

DOI: https://doi.org/10.18502/fbt.v10i2.12216

Keywords: Brain Image Registration; Clustered Redundant Keypoint Elimination Method - Scale Invariant Feature Transform; Spatial Relations; Redundant Keypoints; Magnetic Resonance Imaging.

Abstract

Purpose: The process of Magnetic Resonance Imaging (MRI) image registration is one of the important branches in MRI image analysis, which is a necessary pre-processing to use the information in these images. The purpose of this paper is to present a new approach for MRI image registration that can maintain the total number of initial matches and have the highest precision.

Materials and Methods: The Clustered Redundant Keypoint Elimination Method-Scale Invariant Feature Transform (CRKEM-SIFT) algorithm has recently been introduced to eliminate redundancies and upgrade the correspondence precision. The disadvantages of this algorithm include the high execution time and the number of incorrect correspondences. In this paper, to increase the accuracy and speed of MRI image registration, the CRKEM method is first used over the Speeded Up Robust Features (SURF) algorithm. Then, Spatial Relations Correspondence (SRC) and Alpha-Trimmed Spatial Relations Correspondence (ATSRC) methods are suggested to improve correspondences. These suggested methods, unlike conventional methods such as Random Sample Consensus (RANSAC(, which only eliminates incorrect correspondences, detect incorrect correspondences based on spatial relationships and turn them into correct correspondences. Converting incorrect correspondences to correct ones can increase the number of correct correspondences and ultimately increase the precision of correspondences.

Results: The simulation results show that the suggested CRKEMSURF-ATSRC approach improves the mean by 28.92% in terms of precision and 37.58% in SITMMC compared to those of the SIFT-ARANSAC method.

Conclusion: The suggested SRC and ATSRC methods use the spatial relations of the initial correspondences to convert the incorrect correspondences into correct ones. The number of initial correspondences is maintained in these suggested approaches. These methods are better than other methods of improving correspondences such as RANSAC, and Graph Transformation Matching (GTM). These suggested methods can be used as a new and efficient approach to improve the correspondence of medical images.