Semi-automated segmentation of the lateral periventricular regions using diffusion magnetic resonance imaging

Graphical abstract

Segmentation of the Lateral Ventricular Perimeter (LVP), on an axial mean diffusivity map of an infant with postheorrhgaic hydrocephalus (PHH). A MATLAB-based script segments a homogenous LVP layer using high spatial resolution parameters (voxel size 1.2 × 1.2 × 1.2 mm3) to perform diffusion MRI analysis of the innermost layer of the LVP. The LVP, which comprises the ventricular zone (VZ), subventricular zone and the subjacent periventricular white matter (PVWM) is critical for forming a brain-CSF barrier and global cortical development (. In PHH, and other neurological conditions the LVP is disrupted and has been associated with severe neuromotor and cognitive impairment. Modified from Rodriguez et. al. 2012. Biol Res 45: 231-241

Abstract

The lateral ventricular perimeter (LVP) of the brain is a critical region because in addition to housing neural stem cells required for brain development, it facilitates cerebrospinal fluid (CSF) bulk flow and functions as a blood-CSF barrier to protect periventricular white matter (PVWM) and other adjacent regions from injurious toxins. LVP injury is common, particularly among preterm infants who sustain intraventricular hemorrhage or post hemorrhagic hydrocephalus and has been associated with poor neurological outcomes. Assessment of the LVP with diffusion MRI has been challenging, primarily due to issues with partial volume artifacts since the LVP region is in close proximity to CSF and other structures of varying signal intensities that may be inadvertently included in LVP segmentation.

This research method presents:

• •A novel MATLAB-based method to segment a homogenous LVP layer using high spatial resolution parameters (voxel size 1.2 × 1.2 × 1.2 mm³) to only capture the innermost layer of the LVP.
• •The segmented LVP is averaged from three contiguous axial slices to increase signal to noise ratio and reduce the effect of any residual volume averaging effect and eliminates manual and inter/intrarater-related errors.

Specifications table

Subject Area	Neuroscience
Method name:	Segmentation of lateral ventricular perimeter regions of interest
More specific subject area	Image processing for brain diffusion MRI
Name and reference of original method	N/A.
Resource availability	1.MATLAB (https://www.mathworks.com/products/matlab.html) 2.The tilefigs.m program, previously described by Charles Plum [2] was used only to display images already created by the algorithm (https://www.mathworks.com/matlabcentral/fileexchange/328-tilefigs-m.

Method details

The semi-automated algorithm for selecting periventricular ROIs begins by reading standard processed DTI files into MATLAB. Images showing mean diffusivity (MD) are displayed sequentially to allow the user to select slices centered on the foramen of Monro. Once this slice is determined, the algorithm automatically selects three total axial slices (one rostral and one caudal to the center slice) for further analysis. The MD image is used since it provides a sharp demarcation between the brain parenchyma and the CSF.

The script then displays a histogram of the MD values present in the center slice, which appears as a bimodal distribution. The lower peak of this distribution represents MD values in the brain parenchyma, and the higher peak represents the cerebrospinal fluid spaces. Based on this histogram, the user determines a threshold value that best separates parenchyma and CSF MD values. Following this, the script produces an image showing voxels with MD values above the threshold. When properly executed, these steps result in a few predominant regions representing the lateral and third ventricles, along with some small peripheral regions of CSF. Through sequential erode and dilate steps, the selected regions are smoothed, and small extraneous spaces are removed.

Next, the script calculates the number of objects in each slice and the area of each object (blob analysis) and displays this information. A predetermined minimum size (we used 10 voxels) is applied to display only areas that are likely to be involved in the ventricular system. The script also determines the center of mass of each object, with the assumption that true ventricular objects will be close to the center of the brain. At this stage, the user selects the number of objects to be further processed, which should include all objects that represent ventricular spaces. The script produces an image of all objects that are greater than the minimum size, and the user is able to click near the center of mass of each object to be included in the final selection. Once the relevant objects are selected, the script saves those selections and removes extraneous objects from further analysis. This process is repeated for each of the three contiguous slices.

After the ventricular spaces have been selected, the script isolates the periventricular voxels. This is done by performing two sequential dilation steps, incorporating voxels that are one and two spaces away from the ventricles and then subtracting the original ventricle objects from the dilated objects. The final periventricular ROI is then saved, and it can be sampled for individual DTI parameters.

The anterior (frontal) and posterior (occipital) horns of the lateral ventricle may be independently segmented. In order to select only the frontal and occipital horns, the previously segmented “full” periventricular objects are further modified as follows.

• •The center point of the periventricular ROIs in the anterior-posterior (A-P) direction is calculated. This divides the brain into an anterior and posterior segment.
• •The script then locates the most anterior and most posterior voxels that are also included in the ROI. This identifies the most anterior and posterior aspects of the ventricles which are then used to select all voxels that are within a predetermined A-P distance from these most extreme points for inclusion in the frontal and occipital horn perimeters.
• •An additional step separates the left frontal, right frontal, left occipital, and right occipital objects for individual analysis.

Step-by-step guide

All MATLAB scripts required to run the program is included here. To run the program effectively, and to avoid bugs, we recommend the folder/file structure outlined on Supplementary Fig. 1. Should the user decide to use our recommended structure, all the dMRI images to be analyzed may be placed into folders separated by their “Study Groups”. The following MATLAB scripts can then be used as is to create the 12 files in the “Codes” folder, which are all required for parsing the LVP (“full’) and frontal-occipital horn (“corners”) regions.

Data validation

The methods presented herein have been validated in a prospective cohort of human infants who were born preterm and sustained intraventricular hemorrhage with or without Posthemorrhagic hydrocephalus, in comparison to preterm infants with no identifiable brain injury and full term healthy infants [1]. To assess for potential differences between the algorithm and manually placed ROIs, dMRI of 20 neonates of varying ventricular morphologies were selected and manual ROIs were placed by an expert rater. Pearson correlation analyses between dMRI measures from both approaches indicated there was a strong positive association between the automated and manual segmentations fractional anisotropy (r(20) = .91, p <.001) and MD (r(20) = .89, p <.001) measures.

Conclusion

We describe a novel semi-automated segmentation algorithm that is able to generate ROIs isolating the perimeter of the lateral ventricles in diffusion-based brain MRI. The algorithm has robust fidelity for locating periventricular regions compared to manually produced ROIs and performs well in various situations of high-grade brain injury and ventriculomegaly. An additional step enables isolation of frontal and occipital regions of the periventricular ROI for separate analysis. Importantly, this method is able to extract regions associated with the ventricular system from other CSF spaces (cysts, cisterns, subarachnoid spaces), which has not been previously possible using traditional approaches.