Comparison of MRI normalization strategies for estimating lesion load in multiple sclerosis
Poster No:
1837
Submission Type:
Abstract Submission
Authors:
Brigitta Malagurski Tortei1, Hibba Yousef1
Institutions:
1Technology Innovation Institute, Biotechnology Research Center, P.O.Box: 9639, Masdar City, Abu Dhabi, United Arab Emirates
First Author:
Brigitta Malagurski Tortei
Technology Innovation Institute, Biotechnology Research Center, P.O.Box: 9639, Masdar City
Abu Dhabi, United Arab Emirates
Technology Innovation Institute, Biotechnology Research Center, P.O.Box: 9639, Masdar City
Abu Dhabi, United Arab Emirates
Co-Author:
Hibba Yousef
Technology Innovation Institute, Biotechnology Research Center, P.O.Box: 9639, Masdar City
Abu Dhabi, United Arab Emirates
Technology Innovation Institute, Biotechnology Research Center, P.O.Box: 9639, Masdar City
Abu Dhabi, United Arab Emirates
Introduction:
The presence of pathological lesions on structural MRI poses a significant challenge for image registration to standard anatomical spaces [1], which is an essential step for comparison of homologous brain regions in group studies. Multiple sclerosis (MS) is one such pathology, characterized by demyelinated lesions throughout the central nervous system [2]. During registration, geometric distortions are introduced, and erroneous removal or inflation of some lesions may occur [3]. Furthermore, the discrepancies between lesioned brains and the target template can substantially reduce registration accuracy [4]. This is particularly relevant in lesion-symptom mapping, that is the calculation of the lesion load that reflects the size of the damage and its overlap with a particular region of interest (gray matter or white matter tracts). Thus, our aim was to compare the effects of four registration pipelines with different mitigation strategies and quantify their effect on lesion load in MS.
Methods:
A publicly available dataset of 30 MS patients was used for this study, consisting of structural MRI modalities (T1w, T2w, FLAIR) and corresponding lesion masks, which were segmented semi-automatically [5]. For each patient, the T1 images were denoised, bias corrected, brain-extracted (HD-BET; [6]) and normalized to MNI152 space using four different ANTs-based nonlinear (3-stage registration: rigid, affine and SyN) processing pipelines [7]. In the first one, the brain images were registered to MNI space using the default ANTs workflow (Figure 1 – NoMask). In the second one, the inverse of the lesion mask was generated and used during the registration to limit the areas of the mapping within an excluded region (Figure 1 – CFM). In the third and the fourth workflow, we used the SynthSR [8] and the VBG [9] tools, respectively, to generate lesion-free T1-weighted images, which should support further image processing in cases where lesions would interfere with the quality of the output.
Lesion load (LL) was calculated using the high-resolution sensorimotor area tract template (SMATT; [10]), which contains 12 bilateral tracts derived from probabilistic tractography based on 6 cortical regions in the primary motor cortex (M1), dorsal premotor cortex (PMd), ventral premotor cortex, supplementary motor area (SMA), pre-supplementary motor area (preSMA), and primary somatosensory cortex. This template was chosen due to the prevalence of MS lesions in white matter and motor disability patterns commonly found in MS patients. Lesion load was defined as the number of lesioned voxels intersecting with a given tract divided by the number of voxels in the tract. The comparison between different registration pipelines and their impact on the SMATT lesion load was evaluated using the Wilcoxon signed-rank test. False discovery rate (FDR) was applied to adjust for multiple comparisons.
Lesion load (LL) was calculated using the high-resolution sensorimotor area tract template (SMATT; [10]), which contains 12 bilateral tracts derived from probabilistic tractography based on 6 cortical regions in the primary motor cortex (M1), dorsal premotor cortex (PMd), ventral premotor cortex, supplementary motor area (SMA), pre-supplementary motor area (preSMA), and primary somatosensory cortex. This template was chosen due to the prevalence of MS lesions in white matter and motor disability patterns commonly found in MS patients. Lesion load was defined as the number of lesioned voxels intersecting with a given tract divided by the number of voxels in the tract. The comparison between different registration pipelines and their impact on the SMATT lesion load was evaluated using the Wilcoxon signed-rank test. False discovery rate (FDR) was applied to adjust for multiple comparisons.
Results:
Nonparametric paired comparisons were conducted between SMATT lesion loads calculated using four different processing workflows. We did not find any statistically significant differences between any of the registration pipelines (FDR p values > 0.05; Figure 1).
Conclusions:
In this preliminary study we showed that the lesion filling and lesion masking, applied during registration to MNI space, didn't have a statistically significant impact on the lesion load calculation within a white matter tract-based template in a small sample of MS subjects. However, further studies should also consider the accuracy of each of these normalization methods, the impact of lesion size and heterogeneity, and the choice of different brain atlases and pathologies (e.g., stroke) on the dis/advantages of each of the preprocessing approaches shown here.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 2
Modeling and Analysis Methods:
Image Registration and Computational Anatomy 1
Methods Development
Keywords:
MRI
Other - Registration, Lesion Load, Multiple Sclerosis
1|2Indicates the priority used for review
·Figure 1. Lesion load according to normalization pipeline
Provide references using author date format
[2] Lassmann, H. (2018). Multiple sclerosis pathology. In Cold Spring Harbor Perspectives in Medicine (Vol. 8, Issue 3). Cold Spring Harbor Laboratory Press. https://doi.org/10.1101/cshperspect.a028936
[3] Andresen, J., Uzunova, H., Ehrhardt, J., Kepp, T., & Handels, H. (2022). Image registration and appearance adaptation in non-correspondent image regions for new MS lesions detection. Frontiers in Neuroscience, 16.
[4] Pappas, I., Hector, H., Haws, K., Curran, B., Kayser, A. S., & D’Esposito, M. (2021). Improved normalization of lesioned brains via cohort-specific templates. Human Brain Mapping, 42(13), 4187–4204.
[5] Lesjak, Ž., Galimzianova, A., Koren, A., Lukin, M., Pernuš, F., Likar, B., & Špiclin, Ž. (2018). A Novel Public MR Image Dataset of Multiple Sclerosis Patients With Lesion Segmentations Based on Multi-rater Consensus. Neuroinformatics, 16(1), 51–63.
[6] Isensee, F., Schell, M., Pflueger, I., Brugnara, G., Bonekamp, D., Neuberger, U., Wick, A., Schlemmer, H. P., Heiland, S., Wick, W., Bendszus, M., Maier‐Hein, K. H., & Kickingereder, P. (2019). Automated brain extraction of multisequence MRI using artificial neural networks. Human Brain Mapping, 40(17), 4952–4964.
[7] Avants, B. B., Tustison, N. J., Song, G., Cook, P. A., Klein, A., & Gee, J. C. (2011). A reproducible evaluation of ANTs similarity metric performance in brain image registration. NeuroImage, 54(3), 2033–2044.
[8] Iglesias, J.E, et al. (2023). SynthSR: A public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images for 3D morphometry. Sci. Adv.9, eadd3607.
[9] Radwan, A.M., Emsell, L., Blommaert, J., Zhylka, A., Kovacs, S., Theys, T., Sollmann, N., Dupont, P., Sunaert, S. (2021). Virtual brain grafting: Enabling whole brain parcellation in the presence of large lesions. Neuroimage, 229:117731.
[10] Archer, D.B., Vaillancourt, D.E., Coombes, S.A. (2018). A Template and Probabilistic Atlas of the Human Sensorimotor Tracts using Diffusion MRI. Cereb Cortex, 28(5):1685-1699.