Manual Segmentation of the Human Choroid Plexus Using Brain MRI

Summary

Despite the crucial role of the choroid plexus in the brain, neuroimaging studies of this structure are scarce due to the lack of reliable automated segmentation tools. The present protocol aims to ensure gold-standard manual segmentation of the choroid plexus that can inform future neuroimaging studies.

Abstract

The choroid plexus has been implicated in neurodevelopment and a range of brain disorders. Evidence demonstrates that the choroid plexus is critical for brain maturation, immune/ inflammatory regulation, and behavioral/cognitive functioning. However, current automated neuroimaging segmentation tools are poor at accurately and reliably segmenting the lateral ventricle choroid plexus. Furthermore, there is no existing tool that segments the choroid plexus located in the third and fourth ventricles of the brain. Thus, a protocol delineating how to segment the choroid plexus in the lateral, third, and fourth ventricle is needed to increase the reliability and replicability of studies examining the choroid plexus in neurodevelopmental and brain disorders. This protocol provides detailed steps to create separately labeled files in 3D Slicer for the choroid plexus based on DICOM or NIFTI images. The choroid plexus will be manually segmented using the axial, sagittal, and coronal planes of T1w images making sure to exclude voxels from gray or white matter structures bordering the ventricles. Windowing will be adjusted to assist in the localization of the choroid plexus and its anatomical boundaries. Methods for assessing accuracy and reliability will be demonstrated as part of this protocol. Gold standard segmentation of the choroid plexus using manual delineations can be used to develop better and more reliable automated segmentation tools that can be openly shared to elucidate changes in the choroid plexus across the lifespan and within various brain disorders.

Introduction

Choroid plexus function
The choroid plexus is a highly vascularized structure in the brain consisting of fenestrated capillaries and a monolayer of choroid plexus epithelial cells¹. The choroid plexus projects into the lateral, third, and fourth cerebral ventricles and produces cerebrospinal fluid (CSF), which plays an important role in neural patterning² and brain physiology^3,4. The choroid plexus secretes neurovascular substances, encompasses a stem-cell like repository, and acts as a physical barrier to impede the entrance of toxic metabolites, an enzymatic barrier to remove moieties that circumvent the physical barrier, and an immunological barrier to protect against foreign invaders⁵. The choroid plexus modulates neurogenesis⁶, synaptic plasticity⁷, inflammation⁸, circadian rhythm^9,10, gut brain-axis¹¹, and cognition¹². Moreover, peripheral cytokines, stress, and infection (including SARS-CoV-2) can disrupt the blood-CSF barrier^13,14,15,16. Thus, the choroid plexus-CSF system is integral for neurodevelopment, neurocircuit maturation, brain homeostasis, and repair¹⁷. Since immune, inflammatory, metabolic, and enzymatic alterations impact the brain, researchers are using neuroimaging tools to assess the role of the choroid plexus across the lifespan and in brain disorders^18,19,20. However, limitations exist in commonly used automated tools for choroid plexus segmentation, such as FreeSurfer, which result in the choroid plexus being poorly segmented. Thus, there is a critical need for ground truth manual segmentation of the choroid plexus that can be used to develop an accurate automated tool for choroid plexus segmentation.

Choroid plexus in neurodevelopment and brain disorders
The role of the choroid plexus in brain disorders has long been neglected, mainly because it was regarded as a supporting player whose role was to cushion the brain and maintain a proper salt balance^2,21. However, the choroid plexus has gained attention as a structure linked to brain disorders such as pain syndromes²², SARS-CoV-2^16,23,24, neurodevelopmental², and brain disorders¹⁹, suggesting a transdiagnostic effect in the development of behavioral disorders. In neurodevelopmental disorders, choroid plexus cysts were associated with an increased risk of developmental delay, attention-deficit/hyperactivity disorder (ADHD), or autism spectrum disorder (ASD)^25,26. Additionally, lateral ventricle choroid plexus volume was found to be increased in patients with ASD²⁷. In brain disorders,choroid plexus abnormalities have been described since 1921 in psychotic disorders^28,29. Previous studies have identified choroid plexus enlargement using FreeSurfer segmentation in a large sample of patients with psychotic disorders compared to both their first-degree relatives and controls¹⁹. These findings were replicated using manually segmented choroid plexus volume in a large sample of clinical high-risk for psychosis population and found that these patients had larger choroid plexus volume compared to healthy controls³⁰. There are a growing number of studies demonstrating choroid plexus enlargement in complex regional pain syndrome²², stroke³¹, multiple sclerosis^20,32, Alzheimer's^33,34, and depression³⁵, with some demonstrating a link between peripheral and brain immune/inflammatory activity. These neuroimaging studies are promising; however, poor lateral ventricle choroid plexus segmentation by FreeSurfer²¹ limits the trustworthiness of automated choroid plexus volume estimation. As a result, studies in multiple sclerosis^20,32, depression³⁵, Alzheimer's³⁴, and early psychosis³⁶ have begun manually segmenting the lateral ventricle choroid plexus, but there are no current guidelines for how to do this, nor is their guidance on segmenting the third and fourth ventricle choroid plexus.

Common segmentation tools exclude the choroid plexus
Brain segmentation pipelines such as FreeSurfer^37,38,39, FMRIB Software Library (FSL)⁴⁰, SLANT⁴¹, and FastSurfer (developed by the co-author Martin Reuter)^42,43, accurately and reliably segment cortical and subcortical structures employing atlas-based (FSL), atlas- and surface-based (FreeSurfer), and deep learning segmentation paradigms (SLANT and FastSurfer). Weaknesses of some of these approaches include processing speed, limited generalization to different scanners, field strengths and voxel sizes^37,44, and forced alignment of the label map in a standard atlas space. However, the capability to segment the choroid plexus and the compatibility with high-resolution MRI is only addressed by FreeSurfer and FastSurfer. The neural networks behind FastSurfer, are trained on FreeSurfer choroid plexus labels, so they inherit FreeSurfer's previously discussed reliability and coverage limitations, with the third and fourth ventricles being ignored²¹. Current limitations for high-resolution MRI also exist, but FreeSurfer's high-resolution stream⁴⁵ and FastSurferVINN⁴³ can be used to handle this issue.

Current choroid plexus segmentation tools
There is only one freely available segmentation tool for the choroid plexus, but segmentation accuracy is limited. Accurate choroid plexus segmentation can be impacted by a variety of factors, including (1) variability in choroid plexus location (spatially non-stationary) due to its location within the ventricles, (2) differences in voxel intensity, contrast, resolution (within-structure heterogeneity) due to cellular heterogeneity, dynamic choroid plexus function, pathological changes, or partial volume effects, (3) age- or pathology-related ventricular size differences impacting choroid plexus size, and (4) proximity to adjacent subcortical structures (hippocampus, amygdala, caudate, and cerebellum), which are also difficult to segment. Given these challenges, FreeSurfer segmentations often under or over-estimate, mislabel or ignore the choroid plexus.

Three recent publications addressed the gap of reliable choroid plexus segmentation with a Gaussian Mixture Model (GMM)⁴⁶, an Axial-MLP⁴⁷, and U-Net-based deep learning approaches⁴⁸. Each model was trained and evaluated using private, manually labeled datasets of at most 150 subjects with a limited diversity of scanners, sites, demographics, and disorders. While these publications^46,48,49 achieved significant improvements over FreeSurfer's choroid plexus segmentation - sometimes doubling the intersection of prediction and ground truth, neither method is (1) validated in high-resolution MRI, (2) has dedicated generalization and reliability analyses, (3) features large representative training and testing datasets, (4) specifically addresses or analyzes choroid plexus segmentation challenges such as partial volume effects, or (5) is publicly available as a ready-to-use tool. Thus, the current "gold standard" for choroid plexus segmentation is manual tracing, e.g., using 3D Slicer⁵⁰ or ITK-SNAP⁵¹, which has not been previously described and has been a major challenge for researchers wishing to examine the role of the choroid plexus in their studies. 3D Slicer was chosen for manual segmentation due to the author's familiarity with the software and because it provides the user with various tools based on different approaches that can be combined to obtain the desired result. Other tools can be used, such as ITK-SNAP, which is primarily oriented on image segmentation, and once the tool is mastered, good results can be obtained by the user. Additionally, the authors have conducted a case-control study demonstrating the high accuracy and reliability of their manual segmentation technique using 3D Slicer³⁰, and that specific methodology is described herein.

Protocol

The present protocol was approved by the Institutional Review Board at Beth Israel Deaconess Medical Center. A healthy subject with a brain MRI scan that was free of artifacts or movement was used for this protocol demonstration, and written informed consent was obtained. A 3.0 T MRI scanner with a 32-channel head coil (see Table of Materials) was used to acquire 3D-T1 images with a 1 mm x 1 mm x 1.2 mm resolution. The MP-RAGE ASSET sequence with a 256 x 256 field of view, TR/TE/TI=7.38/3.06/400 ms, and an 11-degree flip angle was used.

1. Importing brain MRI to 3D Slicer

NOTE: 3D Slicer provides documentation related to its user interface.

1. Prepare the brain MRI DICOM (Digital Imaging and Communications in Medicine) or NIFTI (Neuroimaging Informatics Technology Initiative) files for importing into 3D Slicer.
2. Import DICOM data by clicking the DCM button in the upper left corner of the toolbar. Then click the Import DICOM files button to import DICOM format data.
3. If the MRI data is in the NIFTI format, import it by clicking the DATA button in the upper left corner of the toolbar. In the pop-up dialog, select Choose directory to Add to batch import NIFTI data in a folder or select Choose File(s) to Add to import specific NIFTI files. Then, click the OK button to proceed with data upload to 3D Slicer.
4. After importation, MRI data will appear in the window on the right displaying axial, sagittal, and coronal planes.
5. Change the layout of the windows by navigating to Layouts and selecting a specific layout. This can be found by either clicking on the Layouts module image in 3D Slicer's toolbar itself or via View > Layouts from the application menu.

2. Downloading DICOM from sample data in 3D Slicer

1. Click the Download sample data button in the opening screen of Welcome to Slicer section. Then select the MRHead button, and it will start the download process, which may take a few minutes.
2. Ensure that the brain MRI data with axial, sagittal, and coronal planes are shown on the right window.

3. Quality control and adjusting the MRI image

1. Determine qualitatively the image quality and the presence of artifacts due to head motion or scanning issues by critically reviewing each MRI slice.
2. Zoom the image slice by right-clicking and moving the mouse up or down to zoom out or in, respectively.
3. To move the image slice around, left-click the image, press and hold the Shift key, and drag the mouse around.
4. Adjusting the image brightness can help with viewing the choroid plexus. To do this, either click on Adjust window/level of volume in the toolbar or left-click the image and move the mouse up or down to raise or lower the brightness, respectively.
5. Adjusting contrast can additionally help with finding the choroid plexus. Left-click on the image slice and move the mouse to the left or right to increase or decrease the contrast, respectively. To determine the appropriate contrast for the choroid plexus, use the deep gray matter nuclei (central masses of gray matter arrayed around the lateral and third ventricles) or the signal intensity shown in the contrast scale bar.
6. Once the preferred contrast is selected, maintain the same contrast throughout segmentation and do not adjust for potential variations in supra- and infra-tentorial regions.

4. Creating the manual segments of the choroid plexus

1. To begin segmentation of the lateral, third, and fourth ventricle choroid plexus, create the segmentation files in the Segment Editor module. To navigate there, click either on the Segment Editor in the toolbar or go to the Modules: drop-down menu and select Segment Editor.
2. Click on the drop-down menu for segmentation to select different segmentations (if multiple segmentations were created) and rename the currently selected segmentation.
3. Use the Master Volume drop-down to choose which NIFTI or DICOM sets need editing. Only when the volume file is selected, the user can start segmenting/editing.
4. Click the Add button twice to add two segments for the lateral ventricle choroid plexus. To rename these, double-click on the name and change them to Right lateral ventricle choroid plexus and Left lateral ventricle choroid plexus.
5. Click the Add button again to add segments for the third and fourth ventricle choroid plexus, and rename them to "3^rd ventricle choroid plexus" and "4^th ventricle choroid plexus".

5. Viewing different slices and segmentations

1. Before editing, perform a background study to know how to move between layouts in the viewing window and how to change the view or opacity of the segmentations.
2. At the top of the viewing window and to the left of the slice slider, click on the pin icon. This will open a drop-down menu, which may vary depending on the specific layout the window is in.
   ​NOTE: Utilizing different layouts can be helpful when segmenting the choroid plexus since its structure can vary between individuals. For example, the 'Conventional' layout allows the user to simultaneously view all three slices and a 3D view of the scene. Choosing 'Red/Yellow/Green slice only' gives the user a close-up view of the 2D slice to allow for a more precise segmentation of the choroid plexus.

6. Delineating lateral ventricle choroid plexus ROIs

NOTE: Image registration to a template is not necessary for manual segmentation.

1. For the lateral ventricle choroid plexus, begin in the axial plane ensuring that the images are positioned based on the bicommissural line. Then use the trigonum collaterale as a reference point for locating the lateral ventricle choroid plexus.
   1. Once edits have been made in the axial plane, move to the remaining views (sagittal and coronal) to ensure that the manual segmentation of the lateral ventricle choroid plexus is not capturing the surrounding brain parenchyma or CSF.
2. To begin editing, click on the segment to work on, and the segment name will become highlighted.
3. Click on the Paint or Draw tool in the Effects section of Segment Editor to begin manual segmentation.
   NOTE: It is best to start segmenting in one plane (coronal, axial, or sagittal), and after segmentation has been completed in all of the slices, move to other planes to check and refine the manual segmentation. It is suggested that the user starts with the axial or coronal planes as the lateral ventricle choroid plexus is more easily seen in these views.
4. When using the Draw tool, left-click and hold down to draw a contour at the boundary of the lateral ventricle choroid plexus. Once traced, right-click to fill in the drawn-in area.
5. When using the Paint tool, first select the diameter of the brush to be used for painting. A 3% or 5% brush is suggested for a more precise delineation of the choroid plexus, while 10% can be useful for larger selections.
6. For either tool, utilize Paint or Erase to rectify any erroneous delineations by adding or removing selections.
   ​NOTE: Referencing other planes of view can aid in identifying the lateral ventricle choroid plexus structure versus other brain structures, such as surrounding gray matter, the fornix, corpus callosum, or hippocampus. The user is encouraged to exclude brain scans that have choroid plexus cysts identified.
7. Use the level of the red nucleus as a landmark for stopping the segmentation of the choroid plexus in the lateral ventricles.

7. Delineating third and fourth ventricle choroid plexus ROIs

NOTE: Higher resolution T1w images (such as 0.7 or 0.8 mm) and those obtained on a 7T MRI would provide a more accurate and reliable manual segmentation of the third and fourth ventricle choroid plexus. Segmenting the third and fourth ventricle choroid plexus is more difficult than the lateral ventricle choroid plexus as these regions can be much smaller and with fewer voxels to delineate.

1. For the third ventricle choroid plexus, begin in the sagittal plane and use the foramen of Monro, fornix, corpus callosum, thalamus, and internal cerebral vein as reference points to pinpoint the choroid plexus in the 3^rd ventricle. Moving between slices within the same plane can aid in determining whether a region is the fornix, thalamus, vein, or third ventricle choroid plexus.
   1. Once edits have been made in the sagittal plane, navigate to the remaining views (axial and coronal) to ensure that the manual segmentation of the third ventricle choroid plexus is not selecting the surrounding brain parenchyma or CSF.
2. Similarly, for the fourth ventricle choroid plexus, begin in the sagittal plane and use the superior cerebellar peduncle, pons, and medulla as reference points to pinpoint the choroid plexus in the fourth ventricle. Moving between slices within the same plane can aid in determining whether a region is the cerebellum, cerebellar tonsil, inferior medullary velum or 4^th ventricle choroid plexus.
   1. Once edits have been completed in the sagittal plane, move to the remaining views (axial and coronal) to ensure that the manual segmentation of the fourth ventricle choroid plexus is not selecting surrounding brain parenchyma or CSF.

8. Calculating the volumes of the choroid plexus

1. From the Modules drop-down menu, navigate to Quantification and select Segment Statistics.
2. Under Inputs, select the new segmentation map for quantifying under the Segmentation tool and choose the MRI volume from the Scalar volume. For Output table (under Output), choose the Table option. Once complete, press Apply, and a table containing the volume of the choroid plexus will appear in various units.

9. Saving the segments and volume results

1. Click the Save button in the upper left corner of the toolbar to save generated files.
2. Save the segmentation files as .nrrd (3D slicer file), .nii.gz (NIFTI file) or .tsv (table file).

10. Determining accuracy, performance, and agreement of the segmentation

NOTE: It is recommended to use the MONAI package (see Table of Materials), which describes the Dice Coefficient (DC) and the DeepMind average Surface Distance (avgSD). Details on DC and avgSD are described below. In order to compute these metrics, readers will need to know how to program (e.g., python, read images from disk, re-format the data to the appropriate input arrays for these functions). There is no user-friendly package that includes all these metrics.

1. The DC score is a standard approach to quantify the overlap of two geometric domains. To compute the average DC score between two segmentations, provide two tensors y_pred and y, i.e., multi-frame images with one frame for each binarized label image. Tensors y_pred and y may contain segmentations of two different manual raters, repeated segmentations of the same rater, or automated prediction and manual ground truth.
   1. Use the function monai.metrics.compute_meandice to calculate the mean DC score.
   2. Generate suitable binary label tensors with monai.transforms.post.
      NOTE: The include_background parameter can be set to False to exclude the first category (channel index 0) from the DC computation, which is, by convention, assumed to be a background.
2. Consider the avgSD score less common, and note that the approach can differ as multiple definitions exist for surface distance. For example, use the max distance (also known as Hausdorff distance, highly sensitive to outliers), the mean distance (as described here), and the 95th percentile (highly robust) as frequently used measures.
   1. Use the function compute_average_surface_distance to calculate the avgSD score.
   2. Ensure that this function computes the Average Surface Distance from y_pred to y under the default setting.
   3. In addition, if symmetric = True, ensure that the average symmetric surface distance between these two inputs is returned.
3. Perform the statistical analysis of the DC and avgSD score across multiple cases can be performed by using the robust Wilcoxon signed-rank test for paired analysis.
4. Consider using the Intraclass correlation coefficient (ICC) as another commonly used method to determine whether multiple participants can be rated reliably by different raters. Remember that ICC operates on a set of paired measurements (e.g., the volume) of segmentations and not on the segmentation images directly. To compute ICC, use R software and R Studio (see Table of Materials), which makes the process straightforward.
   1. Download the package using install.packages("psych") and load the library(psych).
   2. Enter the data frame, which includes the participants (rows) and a rater in each column, by using Data <- data.frame(df). Then visualize the measurements using plot (Data).
   3. To run ICC, use ICC(Data), which generates a table of the different types of ICC, e.g., to obtain the inter- or intra-rater scores.

Results

The proposed method has undergone iterative refinement for the lateral ventricle choroid plexus, involving extensive testing on a cohort of 169 healthy controls and 340 patients with clinically high risk for psychosis³⁰. Using the technique described above, the authors obtained high intra-rater accuracy and reliability with a DC = 0.89, avgHD = 3.27 mm³, and single-rater ICC = 0.9730, demonstrating the strength of the protocol described herein.

Handli...

Discussion

Critical steps of the protocol
Three critical steps require special attention when implementing this protocol. First, checking the quality and contrast of MR images is key to ensuring accurate segmentation. If the quality of the image is too poor, or the contrast is too low or too high, it may lead to the inaccurate delineation of the choroid plexus. The contrast for the image can be adjusted by viewing the image's grayscale value or by calibrating the values to enhance the contrast between the...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
3D Slicer	3D Slicer	https://www.slicer.org/	A free, open source software for visualization, processing, segmentation, registration, and analysis of medical, biomedical, and other 3D images and meshes; and planning and navigating image-guided procedures.
FreeSurfer	FreeSurfer	https://surfer.nmr.mgh.harvard.edu/	An open source neuroimaging toolkit for processing, analyzing, and visualizing human brain MR images
ITK-SNAP	ITK-SNAP	http://www.itksnap.org/pmwiki/pmwiki.php	A free, open-source, multi-platform software application used to segment structures in 3D and 4D biomedical images.
Monai Package	Monai Consortium	https://docs.monai.io/en/stable/metrics.html	Use for Dice Coefficient and DeepMind average Surface Distance.
MRI scanner	GE	Discovery MR750
Psych Package	R-Project	https://cran.r-project.org/web/packages/psych/index.html	A general purpose toolbox developed originally for personality, psychometric theory and experimental psychology.
R Software	R-Project	https://www.r-project.org/	R is a free software environment for statistical computing and graphics.
RStudio	Posit	https://posit.co/	An RStudio integrated development environment (IDE) is a set of tools built to help you be more productive with R and Python.
Windows or Apple OS Desktop or Laptop	Any company	n/a	Needed for running the software used in this protocol.