HBM︱Region-based spatial standardization method of brain MRI images to achieve accurate registration of brain regions

Written by ︱ He Hengda Editor ︱ Wang Sizhen In many neuroimaging studies, spatial normalization is an essential step in the process of image data preprocessing
.

For group-level analysis between subjects and between experimental groups, spatial normalization warped each subject's brain MRI images onto a standard space
.

Such a brain image registration process allows all subjects' neuroanatomical structures to correspond to each other in a standard space
.

However, spatial standardization and registration of the brain (especially the cerebral cortex) is difficult because of the intricate gyri and sulcus structures of the human brain and the large morphological differences between subjects
.

Such difficulties not only affect the accuracy of spatial normalization, but also increase the likelihood of false-positive results in neuroimaging studies [1, 2]
.

However, even the current state-of-the-art spatial normalization methods are difficult to accurately register cortical regions, with only 60% to 70% of the region registration similarity [3, 4]
.

 On April 12, 2022, Hengda He from the Department of Biomedical Engineering at Columbia University and Prof.
Qolamreza R.
Razlighi from the Department of Radiology, Weill Medical College, Cornell University collaborated on Human Brain Mapping to publish a paper entitled "Landmark- "Guided region-based spatial normalization for functional magnetic resonance imaging", proposes a novel landmark-guided region-based spatial normalization (LG-RBSN) method based on landmark-guided region-based spatial normalization for functional magnetic resonance imaging.

.

Compared with the current state-of-the-art image registration software ANTs (advanced normalization tools), the LG-RBSN method proposed by the researchers greatly improved the correspondence between the cerebral cortical regions between subjects in the experiment, and improved functional magnetic resonance imaging.
Sensitivity and specificity of group-level analysis of imaging (fMRI)
.

The current mainstream spatial standardization methods for brain functional MRI images include volume-based registration, surface-based registration, and volume and surface hybrid registration.
)
.

Compared with the volume registration method, the LG-RBSN proposed by the researchers does not use the classical registration similarity measure based on image voxel intensity, but adopts a fully automatic feature point extraction method and uses The registration of each corresponding region is guided by feature points, which are taken from the surface model vertices of the white matter, gray matter and cerebrospinal fluid segmentation planes
.

LG-RBSN first used the surface registration method to establish the correspondence between the feature points between the subject brain and the standard brain (see the left of Figure 1), and extended the surface registration results to the three-dimensional Euclidean space, so that The surface mapping process of fMRI volume data in the classical surface registration method is avoided
.

This feature makes LG-RBSN applicable not only to the cerebral cortex, but also to other brain regions such as subcortical and cerebellum
.

The current mainstream spatial normalization methods are all solving the optimization problem of whole-brain registration.
However, the complex morphological structure of the brain and the huge differences in brain morphology between subjects make this optimization problem limited to the local minimum
.

Different from these methods, LG-RBSN adopts a region-based spatial normalization method
.

Compared with the whole brain, the morphological structure of each individual brain region is simpler and the structural differences between subjects are smaller.
It is based on this idea that LG-RBSN registers each brain region independently
.

Referring to Figure 1, in the LG-RBSN method, the researchers propose an inverse distance weighted interpolation method to stitch the local nonlinear displacement field of each brain region obtained by combining the registration.
, resulting in a smooth global nonlinear displacement field applicable to the whole brain
.

In order to preserve the topological structure of the image during deformation, the researchers propose a residual compensation method to ensure the bijectivity of the global deformation field
.

Figure 1.
Feature point and region-based spatial normalization method (LG-RBSN) (Source: He H & Razlighi Q, Human Brain Mapping, 2022) The researchers evaluated the performance of the LG-RBSN method by using simulated and real data
.

In experiments with real data, the researchers first used structural MRI data
.

Figure 2 shows the results of the structural MRI data of three of the subjects before and after spatial normalization.
The results of the LG-RBSN were more similar to the MNI152 standard brain reference image than using the ANTs software
.

In the quantitative assessment, the researchers used FreeSurfer software to perform automated brain segmentation for each subject's structural MRI images [5]
.

These segmented regions will be used to evaluate the accuracy of spatial standardization.
The researchers calculated the binarized image similarity between each subject brain region before and after spatial standardization and the corresponding reference standard brain region (Dice similarity coefficient, value range 0 -1)
.

Compared with other spatial normalization methods, LG-RBSN greatly improved the similarity between cerebral cortical regions and MNI152 standard brain counterparts (0.
8558 ± 0.
0080; mean ± SD)
.

The ANTs registration similarity result was 0.
5115 ± 0.
0641 (mean ± standard deviation), and the CVS (combined volumetric and surface registration) similarity result was 0.
6593 ± 0.
0197 (mean ± standard deviation)
.

Figure 2 Comparison of results of different spatial normalization methods: the first and last columns show the subject’s moving image and the reference image (fixed imaging) of the MNI152 standard brain, respectively; the second and third columns show the subject’s brain, respectively The MRI images were spatially normalized by ANTs and LG-RBSN
.

(Credit: He H & Razlighi Q, Human Brain Mapping, 2022) The researchers also used functional MRI data to evaluate the spatially normalized results of the LG-RBSN
.

In task-based functional magnetic resonance imaging (task fMRI) experiments, the researchers used an event-related design and employed visual and auditory stimuli
.

Receiver operating characteristic (ROC) and area under the curve (AUC) were used to assess the sensitivity and specificity of visual and auditory activation of brain regions in group-level analyses
.

The experimental results show that the use of LG-RBSN can relatively improve the area under the curve of the receiver operating characteristic curve by 6.
3% (compared to ANTs) and 1.
1% (compared to CVS)
.

 The classical spatial normalization method realizes the normalization process from the subject brain to the standard brain by optimizing the algorithm to estimate a non-linear displacement field of the whole brain
.

Compared with these methods, LG-RBSN is the first method to achieve spatial normalization by splicing and assembling nonlinear displacement fields of each region to obtain displacement fields applicable to the whole brain
.

Thanks to the simpler brain structure and registration optimization problem in each region, LG-RBSN can achieve more accurate brain image registration
.

However, the differences between the morphological structure and functional architecture of the brain, how to use fMRI to evaluate the results of spatial normalization, and the comparison between the results of the volume-based registration method and the surface-based registration method, all need to be further studied
.

 Conclusion and discussion, inspiration and prospect In conclusion, the researchers propose a new method of spatial normalization based on feature points and regions (LG-RBSN), which makes the cortical regions of the tested cerebral cortex and their corresponding MNI152 Standard cortical regions are highly corresponding after spatial normalization
.

By experimenting with simulated data, real structural and functional MRI data, our proposed method achieves higher structural registration correspondence to neuroanatomical regions than existing best-performing whole-brain-based methods , and higher sensitivity and specificity for group-level analysis of fMRI
.

However, the proposed LG-RBSN method requires more in-depth effect evaluation, such as the effect on the results of group-level independent components analysis
.

The limited accuracy of traditional spatial normalization methods is particularly prominent when dealing with aging and clinical medical populations
.

For example, in the group-level analysis comparing young and old subjects, the current spatial standardization method with limited accuracy is difficult to eliminate the experimental group bias caused by brain atrophy, which may lead to false positive experimental results and conclusions [2]
.

The effect of age and aging on spatially normalized outcomes also requires further research
.

Link to the original text: https://onlinelibrary.
wiley.
com/doi/full/10.
1002/hbm.
25865 The first author and corresponding author of the article He Hengda (photo provided by the author) He Hengda, a doctoral student in the Department of Biomedical Engineering, Columbia University, in 2017 Graduated from the Department of Optoelectronic Information Science and Engineering, Huazhong University of Science and Technology, 2017-2019, studied functional MRI and medical image registration with Professor Dr.
Qolamreza R.
Razlighi in the Department of Neurology, Columbia University Medical Research Center, 2020 Up to now, in the Intelligent Imaging and Neural Computing Laboratory of Columbia University, he has been researching synchronous EEG-fMRI, transcranial magnetic stimulation and brain functional connection with Prof.
Dr.
Paul Sajda
.

Fund support for this article: NIH-NIA K01 AG044467, R01 AG057962 Talent recruitment [1] "Logical Neuroscience" is looking for an associate editor/editor/operation position (online office) Selected articles from previous issues [1] J Neuroinflammation︱Peng Ying's research group Revealing the regulatory role of microglia mitophagy in morphine-induced central nervous system inflammatory suppression【2】Curr Biol︱The relationship between novelty detection and surprise and recency in the primate brain【3】Neurosci Bull︱Qian Ling Jia's research group revealed that homocysteine ​​affects cognitive function by regulating DNA methylation during chronic stress [4] Front Aging Neurosci︱Ma Tao's team revealed that traditional Chinese medicine compounds can improve Alzheimer's disease through multiple pathways and multiple targets Mechanism of energy metabolism in disease【5】Aging Cell︱Gao Xu’s team found that good sleep quality can delay the accelerated aging caused by air pollution【6】Autophagy︱Shen Hanming’s research group revealed that autophagy-related protein WIPI2 regulates mitochondrial outer membrane protein degradation and mitochondrial The new mechanism of autophagy [7] Neuron's heavy review ︱ Sheng Zuhang's team focused on the important role of axonal mitochondrial maintenance and energy supply in neurodegenerative diseases and post-neural injury repair [8] Cell Death Dis︱ Kong Hui et al.
The role of P2X7/NLRP3 inflammasome pathway in early diabetic retinopathy【9】Sci Adv︱Liu Xingguo/Tian Mei’s team discovered a new mechanism of mitochondrial clearance in drug-induced Parkinson’s syndrome【10】Front Aging Neurosci︱Gut preparation can Affecting postoperative delirium by changing the composition of microflora is recommended for high-quality scientific research training courses [1] Symposium on Patch Clamp and Optogenetics and Calcium Imaging Technology Tencent Conference on May 14-15 [2] Scientific Research Skills︱ The 4th Near Infrared Brain Function Data Analysis Class (Online: 2022.
4.
18~4.
30) References (swipe up and down to read) 1, Desai, R.
, Liebenthal, E.
, Possing, ET, Waldron, E.
, & Binder, JR (2005).
Volumetric vs.
surface-based alignment for localization of auditory cortex activation.
NeuroImage, 26, 1019– 1029.
2, Liu, X.
, Gerraty, RT, Grinband, J.
, Parker, D.
, & Razlighi, QR (2017).
Brain atrophy can introduce age-related differences in BOLD response.
Human Brain Mapping, 38, 3402– 3414.
3, Klein, A.
, Andersson, J.
, Ardekani, BA, Ashburner, J.
, Avants, B.
, Chiang, MC, Christensen, GE, Collins, DL, Gee, J .
, Hellier, P.
, Song, JH, Jenkinson, M.
, Lepage, C.
, Rueckert, D.
, Thompson, P.
, Vercauteren, T.
, Woods, RP, Mann, JJ, & Parsey, RV (2009 ).
Evaluation of 14 nonlinear deformation algorithms applied to human brain MRI registration.
NeuroImage, 46, 786– 802.
4, Balakrishnan, G.
, Zhao, A.
, Sabuncu, MR, Guttag, J.
, & Dalca, AV (2019 ).
VoxelMorph: A learning framework for deformable medical image registration.
IEEE Transactions on Medical Imaging, 38, 1788– 1800.
5, Desikan, RS,Ségonne, F.
, Fischl, B.
, Quinn, BT, Dickerson, BC, Blacker, D.
, Buckner, RL, Dale, AM, Maguire, RP, Hyman, BT, Albert, MS, & Killiany, RJ (2006) .
An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest.
NeuroImage, 31, 968– 980.
Plate making︱Sizhen Wang End of this paper