Validation of a semi-automatic coregistration of MRI scans in brain tumor patients during treatment follow-up
Jiun-Lin Yan1,2,3, Anouk van der Hoorn4,5, Timothy J Larkin6, Natalie R Boonzaier6, Tomasz Matys5, and Stephen J Price6

1Clinical Neuroscience, University of Cambridge, Cambridge, United Kingdom, 2Neurosurgery, Chang Gung Memorial Hospital, Keelung, Taiwan, 3Department of neurosurgery, Chang Gung University College of Medicine, Taoyuan, Taiwan, 4Department of radiology (EB44), University Medical Centre Groningen, Groningen, Netherlands, 5Department of radiology, University of Cambridge, Cambridge, United Kingdom, 6Brain tumour imaging laboratory, University of Cambridge, Cambridge, United Kingdom


Coregistration of lesional brain MRI between different time points is challenging. We aimed to propose a two staged semi-automatic coregistration methods to overcome the difficulty. Firstly, we calculated the transformation between presurgical tumor and postsurgical resection cavity by using the linear FLIRT co-registration. This creates a transformation matrix used for the progression and pseudoprogression area with optimal correction of variable brain shift. Then we applied this transformation matrix to a non-linear FNIRT transformation to coregister the brain. Validation by using registration target error showed smaller deviation can be achieved by using this method compared to direct non-linear registration.


To propose a two stage semi-automatic imaging coregistration pipeline to coregister glioblastoma brain MRI between time points of preoperative, postoperative and progression.


There is an expanding research interest in high grade gliomas due to their significant population burden and poor survival despite the extensive standard multimodal treatment. One of the obstacles is the lack of individualized monitoring of tumor characteristics and treatment response before, during and after treatment.3,4 We developed a semi-automatic method (SAC) to coregister MRI scans before and after surgical and after adjuvant treatment of high grade gliomas.


Coregistration methods Coregistration was performed using a two stage semi-automatic coregistration (SAC) (Figure 1). Conventional contrast-enhanced T1-weighted images were coregistered using tools from the FMRIB Software Library (FSL) version First stage was the coregistration of the binary masks consisting of outer contour of the brain, ventricles, and contrast enhancing area (presurgical MRI images) or resection area (follow up MRI images) of each subject between different time points to create a transformation matrix (using FLIRT function). The contrast enhancing area is targeted for resection and so forms most of the resection cavity on the direct postoperative and later follow-up MRI scans. Therefore, this stage of coregistration allowed for optimal correction of variable brain shift at different time points. Secondly, further applied this transformation matrix as input for a nonlinear transformation matrix of the brain images (using the FNIRT functions). Standard linear coregistration (FLIRT) and standard non-linear coregistration (FNIRT) were done separately for comparison. Validation methods Validation was performed using a targeted registration error method for calculating the error in the x and y direction for the cerebral aqueduct and septum pellucidum on the same axial slice (Figure 2A), and in the y and z direction (Figure 2B) for the upper anterior boundary of the third ventricle at the level of the foramen of Monro (Figure 2C) on the same coronal slice. The central point of the tumor/cavity was also targeted for registration error as the location were expected most errors could be. All targeted registration errors were calculated for the SAC method and compared to FLIRT and FNIRT. In addition, we created a 3D structural similarity map using Matlab to compare images in different time points.


Targeted registration error The targeted registration error showed good performance of the SAC method with a clear benefit over FLIRT and standard FNIRT coregistration (Table 1). In the coregistration of postoperative to preoperative, the SAC showed a smaller deviation of vector of cerebral aqueduct (1.1 versus 1.6, p=0.015), septum pallucidum y coordinate (1.3 versus 2.0, p=0.029) and vector (1.8 versus 2.6, p=0.021), and upper most of third ventricle y, z coordinate and vector (0.4 versus 2.2, p<0.001; 1.2 versus 1.9, p=0.043; 1.3 versus 3.3, p<0.001). The SAC also outperformed the default FNIRT coregistration, small deviation can be seen in most of the coordinate and vector of the error targeted points. This coregistration benefit can also be seen when coregistered recurrent images to preoperative images (Table 2). The cerebral aqueduct and the third ventricle both displayed smaller errors for the SAC coregistration than FLIRT and FNIRT. 3D structural similarity Mean 3D structural similarity showed the relative performance of the coregistration method (Figure 3). The peripheral areas including the frontal and parietal areas demonstrated the best performance. A good performance is also seen at the periventricular regions. A relatively lower overlap between the coregistered and reference preoperative scan was seen in the midsagittal and central areas, the centrum semiovale and central cerebellum.


Most coregistration methods focus on coregistration of different sequences of the same scan time point with linear and nonlinear registration methods in healthy subjects and brain tumor patients. They demonstrate good performances for this intrasubject coregistration of data from the same time1,2 or to a standardized brain atlas.6 We developed and validated a reproducible two stages semi-automatic coregistration method using the FSL which is free available for the coregistration in different time points. By using the semi-automatic derived mask of the outer brain contour, ventricles and lesion allowed an optimal estimation of the changes in brain shift and postoperative changes.


The semi-automatic non-linear coregistration allowed for optimal correction of variable brain shift at different time points as evaluated by minimal targeted registration error. This allows for accurate evaluation of treatment response, essential for the growing research area of brain tumor imaging and treatment response evaluation in large sets of patients.


This research was funded by a National Institute of Health Clinician Scientist Fellowship [SJP], a Remmert Adriaan Laan Fund [AH], a René Vogels Fund [AH] and a grant from the Chang Gung Medical Foundation and Chang Gung Memorial Hospital, Keelung [JLY]. None of the authors have financial of other conflict of interest related to the work presented in this paper.


1. Jenkinson M, Bannister P, Brady M, Smith S: Improved Optimization for the Robust and Accurate Linear Registration and Motion Correction of Brain Images. NeuroImage 17:825-841, 2002

2. Klein A, Andersson J, Ardekani BA, Ashburner J, Avants B, Chiang MC, et al: Evaluation of 14 nonlinear deformation algorithms applied to human brain MRI registration. Neuroimage 46:786-802, 2009

3. Nabavi A, Black PM, Gering DT, Westin CF, Mehta V: Serial Intraoperative MR Imaging of Brain Shift. Neurosurgery 48:787-797, 2001

4. Shibamoto Y, Baba F, Oda K, Hayashi S, Kokubo M, Ishihara S, et al: Incidence of brain atrophy and decline in mini-mental state examination score after whole-brain radiotherapy in patients with brain metastases: a prospective study. Int J Radiat Oncol Biol Phys 72:1168-1173, 2008

5. Smith SM: Fast robust automated brain extraction. Hum Brain Mapp 17:143-155, 2002

6. Zacharaki EI, Hogea CS, Shen D, Biros G, Davatzikos C: Non-diffeomorphic registration of brain tumor images by simulating tissue loss and tumor growth. Neuroimage 46:762-774, 2009


FIGURE 1 – Coregistration steps

The steps for the semi-automatic coregistration of the follow-up image with the preoperative reference image are illustrated. Number of the steps and filenames correspond to the text in the supporting material. The corresponding code can also be found in the text of the supporting material.

Targeted registration errors are shown by septum pellucidum (A), cerebral aqueduct (B) and third ventricle (C). Squares (<□) indicate coordinates of the presurgical, circles (●○) indicate the semi-automatic non-linear coregistered images, and triangles (▲∆) indicate the FNIRT coregistered images. Close and open indicator are used for mean and individual respectively.

The group mean 3D structural similarity is shown for the direct postoperative MRI (A) and later recurrence MRI (B), both coregistered to the preoperative MRI. Values indicate relative structural similarity between the coregistered and original preoperative scan, with higher values indicating a greater similarity between the images being compared.

Targeted registration error are provided with the deviations (mm) of the anatomical landmarks coordinates to the reference images after coregistration of the postoperative to the preoperative scan and the preoperative with the recurrence scan by using the semi-automatic non-linear coregistration (SAC), linear (FLIRT) and default non-linear (FNIRT) coregistration. An * indicates statistical significance.

Targeted registration error are provided with the deviations (mm) of the anatomical landmarks coordinates to the reference images after coregistration of the recurrence to the preoperative scan and the preoperative with the recurrence scan by using the semi-automatic non-linear coregistration (SAC), linear (FLIRT) and default non-linear (FNIRT) coregistration. An * indicates statistical significance

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)