2131

Clustering for Anatomical Quantification and Evaluation (CAQE): Unsupervised brain segmentation with quantitative MRI
Sharada Balaji1, Marek Obajtek1, Irene M. Vavasour1, Adam Dvorak1, Guillaume Gilbert2, Roger Tam1, Cornelia Laule1,3, David K.B. Li1, Anthony Traboulsee1, Alex L. MacKay1, and Shannon H. Kolind1
1University of British Columbia, Vancouver, BC, Canada, 2MR Clinical Science, Philips Healthcare Canada, Missisauga, ON, Canada, 3International Collaboration on Repair Discoveries, Vancouver, BC, Canada

Synopsis

Keywords: Microstructure, Microstructure

Motivation: Traditional image segmentation uses conventional MRI to classify tissue types based on image intensity. However, segmenting using microstructural data may provide more specific classification of tissue.

Goal(s): Cluster and segment brain tissue using quantitative MRI measures, to classify tissue based only on microstructural features without spatial input.

Approach: Measures denoting myelin content, anisotropy and tissue heterogeneity were clustered and used to label test datasets based on microstructural features alone, using an unsupervised approach.

Results: Segmentations were more informative than traditional segmentation, and consistent between healthy subjects. Differences between clusters reflect microstructural feature differences which would otherwise be invisible with conventional imaging.

Impact: The CAQE framework can be used for segmentation of tissue based on quantitative measures alone, providing better delineation of regions based on microstructural features. This allows for future comparisons between healthy and damaged tissue, to visualise and interpret pathological changes.

Introduction

Whole-brain image segmentation is typically performed on conventional MRI images and results in separation of white and grey matter (WM, GM) and cerebrospinal fluid (CSF) based on image intensity. However, regions within WM or GM with specific microstructural features cannot easily be distinguished with this approach. Quantitative MRI measures allow for characterization of tissue microstructure. Myelin water imaging (MWI) provides the myelin water fraction (MWF)1, a measure of myelin content, while tensor-valued diffusion encoding (TVDE) provides µFA and CMD to denote local anisotropy and tissue heterogeneity respectively2. Here we present the Clustering for Anatomical Quantification and Evaluation (CAQE) framework for segmenting brain tissue using a combination of MWF, µFA and CMD without spatial input, resulting in segmentations that are driven only by microstructural features.

Methods

Acquisition
25 healthy volunteers (mean=46y, 11M/14F) at 3T:
- T1w: 3D TFE (1x1x1mm3, ~3.5 minutes)
- MWI: CALIPR3 (56 echoes, acquired at 1.7x1.7x1.7mm3, ~7.5 minutes)
- TVDE: Spherical (33 directions, ~3 minutes), planar (32 directions, ~3 minutes) and linear tensor encoding (59 directions, ~5 minutes) at 3x3x3mm3, bmax=2000s/mm2.

Processing
MWF maps were generated using a 3D spatial correlation based analysis4. Maps of µFA and CMD were generated from TVDE data using the QTI+ framework5. Atlases of MWF, µFA and CMD were generated from all participants’ data6 to provide a population average for future comparisons with disease.

Segmentation
The 25 subjects were randomly divided into a training (n=20) and testing (n=5) set, CSF was masked out7 and metric maps were registered8 to each individual’s MWI space. Data from training subjects was loaded into a single pool where individual datapoints were a specific combination of MWF, µFA, and CMD, with no spatial input involved. This pool was scaled and clustered using a Fuzzy C-Means algorithm9 to label datapoints. Each test subject’s data was then classified based on the clustering using a K-Nearest Neighbours algorithm10. The “best” number of clusters was determined using Calinski-Harabasz and Davies-Bouldin scores10.

MWF, µFA and CMD atlases were separately classified based on training with all subjects, to serve as a population-averaged segmentation for comparison with individual subjects.

The corpus callosum (CC) and thalamus were also masked, clustered and classified separately to see whether the segmentations and associated microstructural features could be compared to known anatomical structure.

Results

Figure 1 shows representative MWF, µFA and CMD metric maps. Figure 2 shows the segmented atlas and a table indicating how each cluster differs relative to average whole-brain values. CAQE resulted in more detailed segmentations than simply grey and white matter. Figure 3 compares the segmented atlas to a representative subject’s segmentation; all subjects showed similar cluster patterns. Clustering only the CC and the thalamus is shown in Figures 4 and 5. Overall, explained variance ratios showed that µFA drove the cluster separations (51%), followed by CMD (32%) and MWF (17%).

Discussion

The CAQE approach of segmenting images based purely on quantitative MRI data provided consistent results across test subjects and atlases. The metrics chosen in this study are known to provide complementary microstructural data11, improving the ability to separate regions with different properties. Atlas segmentations corresponded well with each individual test subject, supporting the utility of the atlas segmentation as a comparative tool.

Clustering specific brain structures shows how the segmentation compares with known anatomical details. The CC has varying fiber diameters and densities; CAQE marks out different portions of the CC, reflecting known anatomy (Figure 4)12. The thalamus has a known structure of sub-nuclei; CAQE provides some of these separations, possibly indicating similar features between the different sub-nuclei (Figure 5)13. Finally, the globus pallidus was consistently segmented in all subjects due to its artificially high MWF caused by high iron concentration14.

Due to the nature of this unsupervised approach and absence of ground truth, drawbacks include an unknown ideal number of clusters; more data can decrease the uncertainty in optimising cluster numbers. The different resolutions of images in this study required interpolation, which may have affected results. Despite this, the ability to reliably segment tissue features based on microstructure alone is promising for studying both anatomical features and changes with disease.

Conclusion

Using MWF, µFA and CMD , the CAQE framework consistently segmented brain tissue providing more microstructural information than simply grey and white matter. Population-averaged atlas segmentations can be used for future comparisons against diseased tissue, to assess classification of tissue damage compared to the healthy average. The CAQE framework can be used with any quantitative MRI metrics for segmentation of any tissue.

Acknowledgements

We thank study participants and the UBC MRI Research Centre’s MR technologists and staff for their time and support. SB is funded by a Natural Sciences and Engineering Research Council (NSERC) Canada Graduate Scholarship- Doctoral; SHK is funded by Canadian Institute for Health Research (CIHR), Multiple Sclerosis Canada, NSERC. GG is an employee of Philips Canada.

References

1. Mackay, A., Whittall, K., Adler, J., Li, D., Paty, D. and Graeb, D. (1994), In vivo visualization of myelin water in brain by magnetic resonance. Magn. Reson. Med., 31: 673-677. https://doi.org/10.1002/mrm.1910310614

2. Westin CF, Knutsson H, Pasternak O, Szczepankiewicz F, Özarslan E, van Westen D, Mattisson C, Bogren M, O'Donnell LJ, Kubicki M, Topgaard D, Nilsson M. Q-space trajectory imaging for multidimensional diffusion MRI of the human brain. Neuroimage. 2016 Jul 15;135:345-62. doi: 10.1016/j.neuroimage.2016.02.039.

3. Dvorak AV, Kumar D, Zhang J, Gilbert G, Balaji S, Wiley N, Laule C, Moore GRW, MacKay AL, Kolind SH. The CALIPR framework for highly accelerated myelin water imaging with improved precision and sensitivity. Sci. Adv. 9,eadh9853(2023).DOI:10.1126/sciadv.adh9853

4. Kumar D, Hariharan H, Faizy TD, Borchert P, Siemonsen S, Fiehler J, Reddy R, Sedlacik J. Using 3D spatial correlations to improve the noise robustness of multi component analysis of 3D multi echo quantitative T2 relaxometry data. NeuroImage 2018;178:583--601. https://doi.org/10.1016/j.neuroimage.2018.05.026

5. Herberthson M, Boito D, Haije TD, Feragen A, Westin CF, Özarslan E. Q-space trajectory imaging with positivity constraints (QTI+). Neuroimage. 2021 Sep;238:118198. doi: 10.1016/j.neuroimage.2021.118198.

6. Avants BB, Yushkevich P, Pluta J, Minkoff D, Korczykowski M, Detre J, Gee JC. The optimal template effect in hippocampus studies of diseased populations. Neuroimage. 2010 Feb 1;49(3):2457-66. doi: 10.1016/j.neuroimage.2009.09.062.

7. Avants BB, Tustison NJ, Wu J, Cook PA, Gee JC. An open source multivariate framework for n-tissue segmentation with evaluation on public data. Neuroinformatics. 2011 Dec;9(4):381-400. doi: 10.1007/s12021-011-9109-y.

8. Avants BB, Epstein CL, Grossman M, Gee JC. Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain. Med Image Anal. 2008 Feb;12(1):26-41. doi: 10.1016/j.media.2007.06.004.

9. Warner J, Sexauer J, scikit-fuzzy, twmeggs, alexsavio, Unnikrishnan A, Castelão G, Pontes FA, Uelwer T, pd2f, laurazh, Batista F, alexbuy, Van den Broeck W, Song W, The Gitter Badger, Pérez RAM, Power JF, Mishra H, … 99991. (2019). JDWarner/scikit-fuzzy: Scikit-Fuzzy version 0.4.2 (v0.4.2). Zenodo. https://doi.org/10.5281/zenodo.3541386

10. Fabian P, Gaël V, Alexandre G, Vincent M, Bertrand T, Olivier G, Mathieu B, Peter P, Ron Weiss, Vincent D, Jake V, Alexandre P, David C, Matthieu B, Matthieu P, Édouard D; JMLR. 12(85):2825−2830, 2011.

11. Balaji S, Dvorak AV, Wiley N, MacMillan EL, Traboulsee A, Vavasour IM, Gilbert G, Moore GRW, Li DKB, Laule C, MacKay AL, Kolind SH. Painting a more complete picture of brain microstructure with myelin water fraction and microscopic fractional anisotropy. Under review, submitted Sep 27 2023.

12. Aboitiz F, Montiel J. One hundred million years of interhemispheric communication: the history of the corpus callosum. Braz J Med Biol Res. 2003 Apr;36(4):409-20. doi: 10.1590/s0100-879x2003000400002.

13. Damiani, D. ; Pereira, L. K. ; Damiani, D. ; Nascimento, A. M. Intelligence neurocircuitry: cortical and subcortical structures. Journal of Morphological Sciences 2017. 34;03:123--129. 10.4322/jms.100417.

14. Pfefferbaum A, Adalsteinsson E, Rohlfing T, Sullivan EV. MRI estimates of brain iron concentration in normal aging: comparison of field-dependent (FDRI) and phase (SWI) methods. Neuroimage. 2009 Aug 15;47(2):493-500. doi: 10.1016/j.neuroimage.2009.05.006.

Figures

Figure 1. Representative maps of MWF, µFA and CMD from a single subject with CSF masked out. The Venn diagram represents how these metrics are related, indicating that they provide largely non-overlapping information. These metrics are therefore good for maximizing the spread of data during the clustering process, resulting in reliable segmentations of structures with different features.

Figure 2. Segmentation of atlases into different clusters, based on training with all 25 subjects. The table indicates the relative average value of each MRI metric within the cluster as high (up arrow), medium (dash) or low (down arrow). Each cluster has a unique combination of metrics, with differences between clusters driven by changes in specific metrics (e.g. #1, #2 are separated by a difference in CMD). #6 represents the globus pallidus, which was separated in all subjects as its own cluster.

Figure 3. Comparison of the segmented atlas and a single subject. Although less “clean”, the single subject shows all the same details as the atlas. In particular, #1-3 appear consistent compared to the atlas, while the grey matter features (#4-6) also show similar patterns.

Figure 4. Clustering and classifying the corpus callosum (CC) alone, presented from the atlas segmentation for clarity. The CC falls into 6 clusters with the splenium, body and genu showing distinct clusters. Note that the edges of the CC (L-R on the axial view) are labelled differently than the centre, and the genu is split into more segments than the splenium particularly at the midline. The arrows in the table indicate relative average values of MWF, µFA and CMD.

Figure 5. Clustering and classifying the thalamus alone, presented from the atlas segmentation for clarity. The thalamus clusters into 3 groups, corresponding roughly to the anterior, lateral and medial nuclei although the shapes are not exact; inexact shapes could indicate common features within the different thalamic nuclei.

Proc. Intl. Soc. Mag. Reson. Med. 32 (2024)
2131
DOI: https://doi.org/10.58530/2024/2131