- Department of Medicine, Federal University of Sergipe, Aracaju, Brazil
- Health Sciences Graduate Program, Federal University of Sergipe, Aracaju, Brazil
Correspondence Address:
Bruno Fernandes de Oliveira Santos, Health Sciences Graduate Program, Federal University of Sergipe, Aracaju, Brazil.
DOI:10.25259/SNI_982_2024
Copyright: © 2025 Surgical Neurology International This is an open-access article distributed under the terms of the Creative Commons Attribution-Non Commercial-Share Alike 4.0 License, which allows others to remix, transform, and build upon the work non-commercially, as long as the author is credited and the new creations are licensed under the identical terms.How to cite this article: Erom Lucas Alves Freitas1, Bruno Fernandes de Oliveira Santos1,2. Artificial intelligence in corticospinal tract segmentation using constrained spherical deconvolution. 31-Jan-2025;16:32
How to cite this URL: Erom Lucas Alves Freitas1, Bruno Fernandes de Oliveira Santos1,2. Artificial intelligence in corticospinal tract segmentation using constrained spherical deconvolution. 31-Jan-2025;16:32. Available from: https://surgicalneurologyint.com/?post_type=surgicalint_articles&p=13353
Abstract
Background: Tractography of cerebral white matter tracts is a technique with applications in neurosurgical planning and the diagnosis of neurological diseases. In this context, the approach based on the constrained spherical deconvolution (CSD) algorithm allows for more efficient and plausible segmentations. This study aimed to compare two CSD techniques for corticospinal tract (CST) segmentation.
Methods: This study examined 40 diffusion-weighted images (DWIs) acquired at 7T from healthy participants in the human connectome project (HCP) and 12 clinical 1.5T DWIs from patients undergoing neurosurgical procedures. Tractography was performed using two techniques: regions of interest-based approach and an automatic approach using the TractSeg neural network. The volume of the CST segmented by the two methods was compared using the Dice similarity coefficient.
Results: There was a low similarity between the CST volumes segmented by the two techniques (Dice index for the HCP: 0.479 ± 0.04; Dice index for the Clinical: 0.404 ± 0.08). However, both techniques achieved high levels of consistency in sequential measurements, with intraclass correlation coefficient values above 0.995 for all comparisons. In addition, all selected metrics showed significant differences when comparing the two techniques (HCP – volume P P = 0.0061, mean diffusivity [MD] P P P = 0.0018, MD P = 0.0018).
Conclusion: Both methods demonstrate a high degree of consistency; however, the automatic approach appears to be more consistent overall. When comparing the CST segmentations between the two methods, we observed only a moderate similarity and differences in all considered metrics.
Keywords: Constrained spherical deconvolution, Corticospinal tract, Tractography probabilistic, TractSeg
INTRODUCTION
The advent of tractography techniques for in vivo segmentation of cerebral white matter tracts has enabled greater precision in neurosurgical planning and expanded the diagnostic arsenal for neurological pathologies.[
Historically, classical DTI tractography faced inherent challenges, such as the issue of crossing fibers and the reconstruction of tracts following more curvilinear trajectories. In this context, the emergence and successful implementation of tractography based on the constrained spherical deconvolution (CSD) algorithm[
The corticospinal tract (CST) stands as the primary descending motor pathway usually explored by neurosurgeons and neuroradiologists.[
The objective of this study was to compare two CSD-based tractography methods using 40 diffusion images from the human connectome project (HCP) of healthy individuals and 12 clinical diffusion-weighted images (DWI) from patients who underwent neurosurgical procedures, with a focus on CST segmentation. The study aimed to evaluate the similarity between the two techniques and the specific consistency of their measurements.
MATERIALS AND METHODS
Image acquisition
HCP dataset
We evaluated a total of 40 DWIs from 7 Tesla magnetic resonance imaging (MRI) scans, randomly selected from anonymized healthy patients in the HCP database, comprising 29 women and 11 men aged between 22 and 35 years. The acquisition protocol for these images included a spin-echo echo-planar imaging sequence with a repetition time (TR) of 7000 ms, TE of 71.2 ms, b-values of 1000 and 2000 s/mm2, and an echo spacing of 0.82 ms. Structural T1-weighted images were obtained using Siemens devices.
Clinical dataset
In addition, we selected 12 DWI datasets from patients with pathological conditions to evaluate the consistency of the two tractography techniques in challenging conditions. Among the selected scans, 10 were from patients with intracranial tumors, and two were from patients with Parkinson’s disease undergoing deep brain stimulation (DBS). This cohort included eight women and four men aged between 6 and 69 years. The cranial MRI images were obtained using a SIGNA Explorer 1.5T scanner (GE Healthcare, Chicago, Illinois, USA), with the following technical specifications: a gradient strength of 40 mT/m, a matrix of 256 × 256 pixels, a field of view of 256 × 256 mm, and a slice thickness of 1 mm. The DWI sequences were acquired in the axial plane with a TR of 8.23 ms and time to echo (TE) of 0.1057 s, using 32 directions. The T1-weighted sequences were acquired in the sagittal plane with a TR of 0.008516 s and TE of 0.003492s.
Inclusion and exclusion criteria for HCP and Clinical data
Inclusion criteria were MRI scans, including anatomical (T1-weighted) images and DTI. Exclusion criteria were poor-quality DWIs or the absence of any required images for processing.
Preprocessing
HCP group
The HCP DWI data had already undergone motion distortion correction using the top-up method and correction for distortions induced by eddy currents; thus, no additional preprocessing steps were necessary.[
Clinical group
The diffusion images of patients selected for this study were subjected to preprocessing steps due to their acquisition in a clinical context. Initially, the tool dwidenoise[
CSD tractography
CSD is a technique that refines another signal-processing method called spherical deconvolution. This method involves recovering a signal based on the convolution of an expected response with a known function.[
Preprocessing for CSD tractography
The next step was to estimate the response function using dwi2response with the Tournier algorithm [
Definition of ROIs
We used the MRtrix3 software package to perform segmentation based on regions of interest (ROIs) defined by atlases. For segmenting the CST, we based our approach on defining cortical and subcortical regions to delineate fiber tracking. Specifically, we selected the precentral gyrus as defined by the Juelich Cytoarchitectonic Atlas,[
The Johns Hopkins University International Consortium for Brain Mapping DTI-81 Atlas (JHU-ICBM-DTI-81)[
The ROIs were registered to each subject’s native diffusion space using diffeomorphic nonlinear co-registration with the assistance of advanced normalization tools (ANTs) software.[
Atlas-based tractography
The probabilistic tractography algorithm chosen was second-order integration over fiber orientation distributions (iFOD2), configured in the tckgen tool of MRtrix3.[
Subsequently, all generated segmentations were registered to the common space of the Montreal Neurological Institute (MNI) fractional anisotropy (FA) map using the functional magnetic resonance imaging of the brain (FMRIB) Software Library[
TractSeg
TractSeg is a tractography tool based on a pretrained convolutional neural network model that automatically segments up to 72 predefined white matter tracts.[
According to the documentation protocol, all diffusion images are aligned to the common MNI space[
In the clinical group, for cases of intracranial tumors, the accuracy of the tractography was intraoperatively evaluated through motor mapping with direct stimulation of the motor area (cortical) and motor fibers (subcortical). The Eximius Med neuronavigation platform (Artis, Brasilia, Brazil) was used for this purpose. The nerve monitoring systems (NIM) Eclipse 32-channel system from Medtronic (Medtronic, Dublin, Ireland) was used in conjunction with the platform Eximius MED. The procedures were performed for DBS cases using stereotaxy Invoked ZD Arc and microstimulation (Inomed Medizintechnik GmbH, Emmendingen, Germany) associated with Eximius Med Software Stereotactic Module (Artis, Brasilia, Brazil).
Metrics
We evaluated the similarity of CST segmentations using MRtrix3 and TractSeg through the Dice-Sørensen coefficient (DICE Index).[
In addition, we selected three metrics for the quantitative evaluation of the tracts: the volume of the density map for each segmentation, the mean FA value, and the mean diffusivity (MD) value along the tract.[
Finally, we extracted the mean values of the metrics for each segmented CST from the patients using the mrstats tool.[
Statistical analysis
We used the DICE Index to evaluate the overlap in CST volume generated by the two techniques. In addition, we performed a consistency analysis of the sequential measurements using the intraclass correlation coefficient (ICC) with a two-way random model. The metrics extracted from each technique were subjected to a Shapiro–Wilk normality test. Continuous variables were compared using the t-test or Wilcoxon test, as appropriate, with a significance level of P < 0.05. Analysis was performed using R statistical software (version 4.3.2) and the PyCharm development environment (March 03, 2023) with Python (version 3.10).
Ethics aspects
The images from the HCP database used in this research are public domain and do not require permission for data use. For the clinical group images, all patients provided informed consent for the use of the images, and all data were anonymized to ensure data privacy. This research was submitted to and approved by the Ethics and Research Committee of the Federal University of Sergipe under process number 6.836.579.
RESULTS
In the clinical group, the mean age was 44.9 years (interquartile range 29.5) for those undergoing brain tumor resection [
When comparing the consistency of the two methods (atlas-based and TractSeg) using the DICE index, they are comparable in the HCP group. However, in clinical cases, the automatic method (TractSeg) demonstrates superior consistency compared to the atlas-based method (0.89 ± 0.01 vs. 0.81 ± 0.04, P < 0.001). Furthermore, TractSeg maintains high consistency regardless of the sample considered (HCP or clinical). In contrast, the consistency of the atlas-based method is significantly influenced by the type of sample (P < 0.001).
The consistency analysis of the two techniques (atlas-based and TractSeg) across consecutive measurements (test-retest) of CST volume, FA, and MD are shown in
Comparisons between the automatic and atlas-based approaches showed significant differences in volume for both the HCP (P < 0.0001) and clinical (P < 0.0001) groups [
The three-dimensional projection of the CST using TractSeg and MRtrix, focusing on the HCP individual with the highest DICE index (0.551), is shown in
Figure 2:
Three-dimensional projection of the corticospinal tract segmentation using the two chosen techniques on clinical T1 images. (a) Corticospinal Tract Segmented through the techniques evaluated in anterior view; (b) corticospinal tract segmented by techniques evaluated in left oblique view; (c) corticospinal Tract Segmented through the techniques evaluated in anterior view; (d) corticospinal tract segmented by techniques evaluated in oblique and left superior view.
In participants who underwent neurosurgical intervention, the segmentation results obtained through TractSeg were consistent with intraoperative monitoring findings. In these cases, the generated tracts, when associated with neuronavigation, aligned with the results of neurophysiological monitoring. Intraoperative neuromonitoring, when available, confirmed the tract localization in the precise location indicated by the tractography in all cases. Only one case was operated on without intraoperative neuromonitoring, making it impossible to confirm the tract localization for that particular case.
DISCUSSION
In this study, we reconstructed the CST using two CSD-based tractography techniques and evaluated the similarity and consistency of the generated segmentations. Our findings demonstrate a high similarity of the CST generated consecutively within each technique, suggesting that both methods produce consistent results. However, the automatic approach appears to be more consistent overall. When comparing the CST segmentations between the two methods, we observed a moderate similarity. Importantly, in terms of clinical relevance, the automatic approach showed consistency with intraoperative findings, validating its potential for surgical planning and guidance.[
Our findings indicate significant differences between the two CSD tractography techniques studied. However, it is noteworthy that both techniques demonstrated high consistency in consecutive segmentations, indicating that, although they produce different results, they are robust in the reproducibility of their outcomes.
In the HCP group, which has a more homogeneous anatomy, the volume generated by TractSeg was more significant than that generated by MRtrix, and this relationship was reversed when segmenting the CST in clinical images with significant structural differences. This paradoxical result may indicate that in the presence of topographical heterogeneity, the automatic approach is more refined and accounts for structural alterations. In contrast, the atlas-based approach is more susceptible to false positives, characterized by the continuation of fiber tracking even when physically impossible.[
Considering the HCP group, the difference in CST generated by each technique is evident, with TractSeg showing better delineation of fibers according to their cortical arrangement. The automatic segmentation better represented curvilinear fibers. In the clinical group, the presence of tumor masses posed a significant challenge for both techniques, and it was clear, as shown in
In the noteworthy case presented in
In patients undergoing neurosurgery, the locations of tracts generated through TractSeg and neuronavigation were consistent with observations in neurophysiological monitoring. Thus, TractSeg is reliable and robust, not only for its high reproducibility but also for its accuracy in neurosurgical applications.[
Our findings point to the superiority of the automatic tractography method through TractSeg compared to the widely used atlas-based approach, which has limitations that compromise its results. In this sense, the automatic segmentation of cerebral white matter tracts, including the CST, represents the state-of-the-art technique in tractography.[
This represents a significant advance for tractography applications in neurosurgical planning and clinical diagnosis of neurological diseases affecting white matter.[
Limitations
The most crucial limitation concerns the implementation of TractSeg, as it is still less common than other approaches and requires programming language knowledge, which can pose execution obstacles. In addition, atlas-assisted ROI demarcation has limitations compared to manual demarcation when considering specific neuroanatomical characteristics.[
The limitation of the clinical image dataset size, although significant, was mitigated by employing robust statistical data analysis techniques aimed at extracting relevant and reliable information.
CONCLUSION
Both methods demonstrate a high degree of consistency; however, the automatic approach appears to be more consistent overall. When comparing the CST segmentations between the two methods, we observed only a moderate similarity and differences in all considered metrics (CST volume, FA, and MD).
Ethical approval
The research/study was approved by the Institutional Review Board at the Ethics and Research Committee of the Federal University of Sergipe, number 6.836.579, dated May 21, 2024.
Declaration of patient consent
The authors certify that they have obtained all appropriate patient consent.
Financial support and sponsorship
Nil.
Conflicts of interest
There are no conflicts of interest.
Use of artificial intelligence (AI)-assisted technology for manuscript preparation:
The authors confirm that there was no use of artificial intelligence (AI)-assisted technology for assisting in the writing or editing of the manuscript and no images were manipulated using AI.
Disclaimer
The views and opinions expressed in this article are those of the authors and do not necessarily reflect the official policy or position of the Journal or its management. The information contained in this article should not be considered to be medical advice; patients should consult their own physicians for advice as to their specific medical needs.
References
1. Amunts K, Mohlberg H, Bludau S, Zilles K. Julich-brain: A 3D probabilistic atlas of the human brain’s cytoarchitecture. Science. 2020. 369: 988-92
2. Andersson JL, Sotiropoulos SN. An integrated approach to correction for off-resonance effects and subject movement in diffusion MR imaging. Neuroimage. 2016. 125: 1063-78
3. Avants BB, Tustison NJ, Song G, Cook PA, Klein A, Gee JC. A reproducible evaluation of ANTs similarity metric performance in brain image registration. Neuroimage. 2011. 54: 2033-44
4. Calamante F, Tournier JD, Heidemann RM, Anwander A, Jackson GD, Connelly A. Track density imaging (TDI): validation of super resolution property. Neuroimage. 2011. 56: 1259-66
5. Fonov V, Evans A, McKinstry R, Almli C, Collins D. Unbiased nonlinear average age-appropriate brain templates from birth to adulthood. Neuroimage. 2009. 47: S102
6. Glasser MF, Smith SM, Marcus DS, Andersson JL, Auerbach EJ, Behrens TE. The Human Connectome Project’s neuroimaging approach. Nat Neurosci. 2016. 19: 1175-87
7. Gutierrez CE, Skibbe H, Nakae K, Tsukada H, Lienard J, Watakabe A. Optimization and validation of diffusion MRI-based fiber tracking with neural tracer data as a reference. Sci Rep. 2020. 10: 21285
8. Hua K, Zhang J, Wakana S, Jiang H, Li X, Reich DS. Tract probability maps in stereotaxic spaces: Analyses of white matter anatomy and tract-specific quantification. Neuroimage. 2008. 39: 336-47
9. Jenkinson M, Bannister P, Brady M, Smith S. Improved optimization for the robust and accurate linear registration and motion correction of brain images. Neuroimage. 2002. 17: 825-41
10. Jenkinson M, Beckmann CF, Behrens TE, Woolrich MW, Smith SM. FSL. Neuroimage. 2012. 62: 782-90
11. Kamagata K, Andica C, Uchida W, Takabayashi K, Saito Y, Lukies M. Advancements in diffusion MRI tractography for neurosurgery. Invest Radiol. 2024. 59: 13-25
12. Kellner E, Dhital B, Kiselev VG, Reisert M. Gibbs-ringing artifact removal based on local subvoxel-shifts. Magn Reson Med. 2016. 76: 1574-81
13. Luque Laguna PA, Combes AJ, Streffer J, Einstein S, Timmers M, Williams SC. Reproducibility, reliability and variability of FA and MD in the older healthy population: A test-retest multiparametric analysis. NeuroImage Clin. 2020. 26: 102168
14. Makris N, Goldstein JM, Kennedy D, Hodge SM, Caviness VS, Faraone SV. Decreased volume of left and total anterior insular lobule in schizophrenia. Schizophr Res. 2006. 83: 155-71
15. Mormina E, Longo M, Arrigo A, Alafaci C, Tomasello F, Calamuneri A. MRI tractography of corticospinal tract and arcuate fasciculus in high-grade gliomas performed by constrained spherical deconvolution: Qualitative and quantitative analysis. Am J Neuroradiol. 2015. 36: 1853-8
16. Moshe YH, Ben Bashat D, Hananis Z, Teicher M, Artzi M. Utilizing the TractSeg tool for automatic corticospinal tract segmentation in patients with brain pathology. Technol Cancer Res Treat. 2022. 21: 15330338221131387
17. O’Donnell LJ, Suter Y, Rigolo L, Kahali P, Zhang F, Norton I. Automated white matter fiber tract identification in patients with brain tumors. NeuroImage Clin. 2017. 13: 138-53
18. Panesar SS, Abhinav K, Yeh FC, Jacquesson T, Collins M, Fernandez-Miranda J. Tractography for surgical neurooncology planning: Towards a gold standard. Neurotherapeutics. 2019. 16: 36-51
19. Petersen MV, McIntyre CC. Comparison of anatomical pathway models with tractography estimates of the pallidothalamic, cerebellothalamic, and corticospinal tracts. Brain Connect. 2023. 13: 237-46
20. Pujol S, Wells W, Pierpaoli C, Brun C, Gee J, Cheng G. The DTI challenge: Toward standardized evaluation of diffusion tensor imaging tractography for neurosurgery. J Neuroimaging. 2015. 25: 875-82
21. Rheault F, De Benedictis A, Daducci A, Maffei C, Tax CM, Romascano D. Tractostorm: The what, why, and how of tractography dissection reproducibility. Hum Brain Mapp. 2020. 41: 1859-74
22. Richards TJ, Anderson KL, Anderson JS. Fully automated segmentation of the corticospinal tract using the TractSeg algorithm in patients with brain tumors. Clin Neurol Neurosurg. 2021. 210: 107001
23. Simon HA. Artificial intelligence: Where has it been, and where is it going?. IEEE Trans Knowl Data Eng. 1991. 3: 128-36
24. Taha AA, Hanbury A. Metrics for evaluating 3D medical image segmentation: Analysis, selection, and tool. BMC Med Imaging. 2015. 15: 29
25. Tournier JD, Calamante F, Connelly A. Determination of the appropriate b value and number of gradient directions for high-angular-resolution diffusion-weighted imaging. NMR Biomed. 2013. 26: 1775-86
26. Tournier JD, Calamante F, Connelly A. MRtrix: Diffusion tractography in crossing fiber regions. Int J Imaging Syst Technol. 2012. 22: 53-66
27. Tournier JD, Calamante F, Connelly A. Robust determination of the fibre orientation distribution in diffusion MRI: Nonnegativity constrained super-resolved spherical deconvolution. Neuroimage. 2007. 35: 1459-72
28. Tournier JD, Smith R, Raffelt D, Tabbara R, Dhollander T, Pietsch M. MRtrix3: A fast, flexible and open software framework for medical image processing and visualisation. Neuroimage. 2019. 202: 116137
29. Veraart J, Novikov DS, Christiaens D, Ades-Aron B, Sijbers J, Fieremans E. Denoising of diffusion MRI using random matrix theory. Neuroimage. 2016. 142: 394-406
30. Wasserthal J, Maier-Hein KH, Neher PF, Northoff G, Kubera KM, Fritze S. Multiparametric mapping of white matter microstructure in catatonia. Neuropsychopharmacology. 2020. 45: 1750-7
31. Wasserthal J, Neher P, Maier-Hein KH. TractSeg-Fast and accurate white matter tract segmentation. Neuroimage. 2018. 183: 239-53
32. Wasserthal J, Neher PF, Hirjak D, Maier-Hein KH. Combined tract segmentation and orientation mapping for bundle-specific tractography. Med Image Anal. 2019. 58: 101559
33. Young F, Aquilina K, Seunarine KK, Mancini L, Clark CA, Clayden JD. Fibre orientation atlas guided rapid segmentation of white matter tracts. Hum Brain Mapp. 2024. 45: e26578
34. Zhylka A, Sollmann N, Kofler F, Radwan A, De Luca A, Gempt J. Reconstruction of the corticospinal tract in patients with motor-eloquent high-grade gliomas using multilevel fiber tractography combined with functional motor cortex mapping. Am J Neuroradiol. 2023. 44: 283-90