An efficient similarity metric for 3D medical image registration

• Published in: Multimedia tools and applications Vol. 83; no. 40; pp. 87987 - 88017
• Main Authors: Sengupta, Debapriya, Gupta, Phalguni, Biswas, Arindam
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.12.2024; Springer Nature B.V
• Subjects: Algorithms; Amoeba; Computation; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Entropy; Health care facilities; Image registration; Image segmentation; Lattice theory; Mattes; Medical imaging; Metric space; Multimedia Information Systems; Registration; Similarity; Special Purpose and Application-Based Systems; Thorax; Track 2: Medical Applications of Multimedia; Entropy; Joint entropy; Mutual information; Medical image registration
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-024-18710-1

In this paper, we develop an efficient mutual information based similarity metric for 3D medical image registration. The efficiency of the metric lies in the computation of mutual information, which uses modified algorithms for calculating entropy and joint entropy. We implemented the newly developed Efficient Entropy Mutual Information (EEMI) metric in SimpleITK, which is an open source image registration and segmentation toolkit. We designed 24 medical image registration frameworks using four similarity metrics, namely Mean Squares (MS), Joint Histogram Mutual Information (JHMI), Mattes Mutual Information (MattesMI) and our proposed EEMI, and six optimizers namely, Gradient Descent (GD), Conjugate Geadient Line Search (CGLS), 1+1 Evolutionary, Powell, Nelder-Mead (Amoeba) and Limited Memory Broyden Fletcher Goldfarb Shannon (LBFGS2). Using these frameworks, we performed elaborate comparative evaluation and analysis of EEMI in terms of registration accuracy and computation time. We used four different medical image data sets for our experiments. Data consisted of intra-modal and inter-modal image pairs of the brain and thorax obtained from different medical institutions as well as publicly available databases. Registration results prove the superiority and consistency of performance of EEMI compared to MS and JHMI. When compared with the benchmark MattesMI metric, performance of EEMI is on a par with respect to Dice score, Jaccard score and Hausdorff distance. With respect to computation time, EEMI is faster with GD, CGLS and Amoeba optimizers with 61.29%, 42.67% and 10.92% gain in computation times respectively. Among the other three optimizers, EEMI is more consistent with Powell’s optimizer. Mean, median and STD of TRE values of EEMI with Powell’s optimizer are respectively 67.42%, 65.84% and 74.63% less than that of MattesMI. Hausdorff distance of EEMI is 47.47% less than MattesMI, with Powell’s optimizer.