Registration of IRT and visible light images in neurosurgery: analysis and comparison of automatic intensity-based registration approaches
Related Experiment Video
Updated: Oct 3, 2025
A Pipeline for 3D Multimodality Image Integration and Computer-assisted Planning in Epilepsy Surgery
Published on: May 20, 2016
Yahya Moshaei-Nezhad1, Juliane Müller2, Martin Oelschlägel2
1Institute of Circuits and Systems, Faculty of Electrical and Computer Engineering, Technische Universität Dresden, 01062, Dresden, Germany. yahya.nezhad@tu-dresden.de.
Purpose:
The purpose of this study is to analyze and compare six automatic intensity-based registration methods for intraoperative infrared thermography (IRT) and visible light imaging (VIS/RGB). The practical requirement is to get a good performance of Euclidean distance between manually set landmarks in reference and target images as well as to achieve a high structural similarity index metric (SSIM) and peak signal-to-noise ratio (PSNR) with respect to the reference image.
Methods:
In this study, preprocessing is applied to bring both image types to a similar intensity. Similarity transformation is employed to align roughly IRT and visible light images. Two optimizers and two measures are used in this process. Thereafter, due to locally different displacement of the brain surface through respiration and heartbeat, two non-rigid transformations are applied, and finally, a bicubic interpolation is carried out to compensate for the resulting estimated transformation. Performance was assessed using eleven image datasets. The registration accuracy of the different computational approaches was assessed based on SSIM and PSNR. Additionally, five concise landmarks for each dataset were selected manually in reference and target images and the Euclidean distance between the corresponding landmarks was compared.
Results:
The results are showing that the combination of normalized intensity, mutual information measure with one-plus-one evolutionary optimizer in combination with Demon registration results in improved accuracy and performance as compared to all other methods tested here. Furthermore, the obtained results led to [Formula: see text], [Formula: see text], [Formula: see text], [Formula: see text], and [Formula: see text] registrations for datasets 1, 2, 5, 7, and 8 with respect to the second best result by calculating the mean Euclidean distance of five landmarks.
Conclusions:
We conclude that the mutual information measure with one-plus-one evolutionary optimizer in combination with Demon registration can achieve better accuracy and performance to those other methods mentioned here for automatic registration of IRT and visible light images in neurosurgery.
