Featured Application: This non-invasive methodology could be used to detect alterations in the cerebrovasculature by analyzing MRA images, which would assist clinicians to optimize medical treatment plans of HBP. Blood pressure (BP) changes with age are widespread, and systemic high blood pressure (HBP) is a serious factor in developing strokes and cognitive impairment. A non-invasive methodology to detect changes in human brain's vasculature using Magnetic Resonance Angiography (MRA) data and correlation of cerebrovascular changes to mean arterial pressure (MAP) is presented. MRA data and systemic blood pressure measurements were gathered from patients (n = 15, M = 8, F = 7, Age = 49.2 ± 7.3 years) over 700 days (an initial visit and then a follow-up period of 2 years with a final visit.). A novel segmentation algorithm was developed to delineate brain blood vessels from surrounding tissue. Vascular probability distribution function (PDF) was calculated from segmentation data to correlate the temporal changes in cerebral vasculature to MAP calculated from systemic BP measurements. A 3D reconstruction of the cerebral vasculature was performed using a growing tree model. Segmentation results recorded 99.9% specificity and 99.7% sensitivity in identifying and delineating the brain's vascular tree. The PDFs had a statistically significant correlation to MAP changes below the circle of Willis (p-value = 0.0007). This non-invasive methodology could be used to detect alterations in the cerebrovascular system by analyzing MRA images, which would assist clinicians in optimizing medical treatment plans of HBP.
Keywords: magnetic resonance angiography (MRA); blood pressure (BP); systolic pressure; diastolic pressure; arteries; cerebral; hypertension
High blood pressure (HBP) affects approximately 1 in 3 adults in the USA. HBP is a primary or a contributing cause of mortality in about 410,000 adults each year with associated healthcare costs of $46 billion [[
Currently, HBP is diagnosed and medically managed when systemic BP measurements using sphygmomanometer are greater than 140/90 mmHg. However, BP measurement via sphygmomanometer cannot quantify cerebrovascular structural changes that can increase the risk of cerebral adverse events. Recently, some studies hypothesized that vascular structural changes in human brains may occur prior to the elevation of systemic BP rather than cerebrovascular damage due to sustained exposure to HBP [[
Magnetic Resonance Imaging (MRI) technique has been traditionally used in the quantification of organ structural changes. In the literature, MRI scanning has been used for volumetric measurement of the ventricular cavities and myocardium [[
The goal of this manuscript is to develop a new framework that detects the potential changes in cerebrovascular structure by first introducing a novel automatic segmentation algorithm that delineates the cerebrovascular system from MRA data, and then estimate the change in cerebral vascular diameters to demonstrate proof-of-concept of correlation between cerebrovascular structural changes to MAP.
In this section, patient demographics and details about the proposed methodology and data analysis are presented.
This work has been approved by the Institutional Review board (IRB) at Pittsburgh University. MRA scans and corresponding blood pressure measurements were acquired from subjects (n = 15, M = 8, F = 7, age = 49.2 ± 7.3) during a 700 day study and were analyzed retrospectively. Participants in this study were carefully selected to represent a wide range of BP changes over the 700 day period and the MRA data were analyzed blinded to the corresponding BP measurements. Participants were chosen in accordance to some exclusion criteria: (
The subjects had an average day 0 systolic pressure of 122 ± 6.9 mmHg, an average day 0 diastolic pressure of 82 ± 3.8 mmHg, at an average day 700 systolic pressure of 118.9 ± 12.4 mmHg, and an average day 700 diastolic pressure of 79.9 ± 11.0 mmHg (i.e., mean systolic pressure remained comparable over time though some individuals increased in pressure and some decreased or stayed the same).
The analysis of patient MRA data consists of five key steps (Figure 1): (
Manual Segmentation of Training Slices: MRA data from a patient consists of 160 MRA slices. Every tenth slice was manually segmented to extract the blood vessels from surrounding tissue using Adobe Photoshop (Adobe Systems, San Jose, CA, USA). This methodology allows for delineation of the in-plane blood vessel from the surrounding tissue at a pixel level accuracy where the largest limitation is the resolution of the MRI machine itself. The manually segmented training binary (black for surrounding tissue and white for target vasculature) slices are referred to as ground truths (GT) as the images are correct and free from artifacts or noise (Figure 2). The manual segmentation of select slices was used for the initialization and optimization of the segmentation algorithm, which was subsequently used for segmenting all obtained slices.
Automatic Segmentation: One of the most challenging issues relating to common computer-assisted diagnostics is the segmentation of accurate 3D cerebrovascular system information from MRA images. Our approach was to rapidly and accurately extract the blood vessel data by defining the probability models for all regions of interest within the statistical approach and not predefining the probability models [[
Adapting the Expectation-Maximization (EM) technique to the LCDG allows for precise identification of the LCDG model which included the number of its components (negative and positive) [[
An expected log-likelihood was used as a criterion for model identification in [[
As explained in [[
A discrete Gaussian distribution is defined on the set of integers (gray levels) Q = {0,1, ..., Q − 1} by the probability mass function:
where the parameter θ = (μ, σ), and Φ is the CDF of a normal distribution with mean μ and variance σ
(
The weights w = (w
(
In general, valid probabilities are nonnegative:
Our aim is finding a K-modal probability model which approximates closely the unknown marginal distribution of the gray level. Consider F
(
The segmentation algorithm basic steps are following [[
- (
1 ) For every slice Xs , s=1,.....S, - (a) First is to gather the marginal empirical probability distribution F
s of gray levels. - (b) Find a starting LCDG model which is nearing F
s by using the initialization algorithm to approximate the values of Cp −K, Cn , and the parameters w, Θ (weights, means, and variances) of the negative and positive discrete Gaussians (DG). - (c) Fixing C
p and Cn , refine the LCDG-model with the modified EM algorithm by manipulating the other parameters.(See Appendix A for more details) - (d) Separate the final LCDG model into K sub models. Each dominant mode has a corresponding sub model. This is done by minimizing the misclassification predicted errors and selecting the LCDG-sub model that has the greatest average value (corresponding to the pixels with highest brightness) to be the model of the wanted vasculature.
- (e) Use intensity threshold t to extract the voxels of the blood vessels in the MRA slice, which separates their LCDG-sub model from the background.
- (
2 ) Remove the artifacts from the extracted voxels whole set with a connection filter which chooses the greatest connected tree system built by a 3D growing algorithm [[23 ]]. Algorithm 1 summarizes the adopted segmentation approach.
LCDG Initialization: Find the marginal empirical probability distribution of gray levels Estimate Find the initial LCDG-model that approximates LCDG Refinement: Fixing Initial Segmentation: Divide the final LCDG-model into K sub models by minimizing the expected errors of misclassification. Select the LCDG-sub model that has the largest mean value to be the model of the wanted vasculature. Use the intensity threshold Final Segmentation: Remove the artifacts from the extracted voxels whole set with a connection filter which chooses the greatest connected tree system built by a 3D growing algorithm.For every slice
This procedure aims to decipher threshold for every MRA image which will enable the complete extraction of the bright blood vessels while removing the darker unwanted tissue and also separating surrounding non-vasculature tissue that may be of similar brightness and along the same boundaries. Step 1b's initialization creates the LCDG with the non-negative starting probabilities p
Accuracy of the automatic segmentation is evaluated by calculating total error compared to the ground truths. True positive (TP), false positive (FP), true negative (TN), and false negative (FN) segmentations are measured for evaluation.
In Figure 3, if C is the segmented region, G is the ground truth, and R represents the entire image frame, then the TP = |C ∩ G|, the TN = |R − C ∪ G|, the FP = |C − C ∩ G|; and the FN = |G − C ∩ G|. The total error ε is given in [[
Voxel Matching: A voxel is an array of volume elements that constitute a notional 3D space. A 3D affine registration is used to handle the pose, orientation, and the data spacing changes and other scanning parameter changes between day 0 and day 700 [[
Generation of Probability Distribution Function and Validation: The EM-based technique is adapted to the LCDG-model and the distribution of pixel distances is extracted from the distance map to calculate the probability distribution of the cerebrovascular changes. The PDF marks the distribution of white pixels as a true value and black pixels being ignored for the data set. The diameters of the blood vessels are determined by estimating Euclidian center point distances from the edge of a vessel. The data points in the generated PDFs are then extracted and compared to the blood pressure data using statistical analysis.
Calculation of Cumulative Distribution Function: The integral of the PDF is used to generate the CDF as follows: CDF (F
Data were statistically analyzed using R-software (version 3.30) by The R Foundation for Statistical Computing, Vienna, Austria. A mixed effects linear model was used to test the relationship of MRA data with clinical BP measurements. Brain slices were separated into upper compartment (above circle of Willis) and lower compartment (below circle of Willis) to determine correlation with clinical BP readings. Circle of Willis, near the brain base, is where the intracranial cerebral arteries take off from and give rise to progressively smaller vessels [[
A growing tree model that eliminates any unwanted segmented voxels by choosing the greatest connected vascular tree system, coupled with a smoothing algorithm, was used to generate a 3-D model based on segmented slices [[
Table 1 shows the specificity and sensitivity of the segmentation algorithm. The automatically segmented slices for all 15 patients were compared to GTs (manually segmented) to calculate accuracy of the algorithm (Figure 4). A cumulative sensitivity of 0.997 ± 0.008 (sensitivity range = 0.969 to 1) and the cumulative specificity of 0.9998 ± 0.0001 (specificity range = 0.9994 to 1) were recorded. The manually segmented training slices (every 10th slice) were excluded in the accuracy calculations of the proposed segmentation approach.
The results of the analysis of the linear mixed effects models (Table 2) revealed that MAP is inversely related to the mean vessel diameter below the circle of Willis (p = 0.0007). The mean diameter of vessels below the circle of Willis was not found to vary significantly with age of the patient or the gender of the patient. Above the circle of Willis, the mean diameters of vessels showed statistically significant decrease with age (p = 0.0005).
In the analysis, 13 out of 15 patients showed significant correlation between MAP and the diameters indicated via CDF. Out of the 13 patients that showed CDF correlation with MAP, two example patients (A and B) are shown with the two patients (C and D) where the correlation between CDF and MAP was not found (Figure 5). Patient C had a shift in CDF in opposition to the MAP change, and patient D had a larger shift in CDF compared to the MAP change (Figure 5C,D, Table 3). The 3D cerebrovascular model reconstruction of patients C and D indicated significant vascular changes between day 0 and day 700 (Figure 6).
The average cumulative segmentation algorithm had a sensitivity of 0.997 ± 0.008 and a specificity of 0.9998 ± 0.001. This high level of accuracy demonstrates the benefit of using a manual input to initialize automatic segmentation. Using manual segmentation alone would be too time intensive to be used in a practical healthcare setting, while utilizing only an automatic segmentation approach would not provide sufficient segmentation accuracy to delineate and quantify the diameters of the smaller arteriolar (<10 micrometers) cerebral blood vessels. The proposed segmentation algorithm combines the accuracy of manual segmentation with the benefit of automated and less time intensive approach and provides segmentation with a high degree of accuracy while also minimizing the required time and effort. Almost all the state-of-the-art algorithms that were used to detect blood vessels were only suitable for healthy blood vessels as they usually assume the vascular linearity and/or circular cross-sections which is not the case in pathological vessels [[
The high degree of sensitivity and specificity of our approach in accurately delineating blood vessels from surrounding brain tissue enables the quantification of cerebrovascular changes. The PDFs indicate the total blood vessel diameter change in time from day 0 and day 700. Below the circle of Willis, there is a statistical correlation between PDFs and systemic BP (p-value = 0.0007), demonstrating that increased MAP is related to decreased average vessel diameter and PDF. The BP and MAP measurements correlate well with most patients' non-invasive mean PDF diameter measurements below the circle of Willis. Since cerebral changes have been hypothesized to precede systemic hypertension [[
In some patients (C and D) the change in CDF did not correlate to changes in MAP, which may indicate impaired auto regulation of cerebral blood flow potentially due to cerebrovascular remodeling [[
The proposed approach uses MRA, which directly visualizes overall cerebrovascular anatomy and provides a higher resolution for small blood vessels compared to Doppler Ultrasound which primarily assesses blood flow or the characteristics of a single or small number of vessels within its field of view. MRA is required for visualization of small cerebral blood vessels to obtain an accurate 3D vascular structure. Routine screening using the proposed MRA-based method would be expensive. The clinical application of importance would be treatment of resistant hypertension, which is now poorly understood [[
The segmentation algorithm and metrics for vascular and blood pressure changes (CDF, PDF) are not limited to cerebral vasculature. These methodologies may also be used to quantify vascular changes in other end organs that are sensitive to blood pressure (e.g., kidneys).
While our segmentation algorithm significantly improves on automatic segmentation methodologies, it is limited by the resolution limit of the MRI machine performing the MRA scanning. The CDF diameters (Figure 5) start at 0.5 mm because the distance map calculations determine radius from the edge of a blood vessel and a pixel in the MRA imaging represented 0.25 mm. Any value less than 0.5 mm would not be accurately represented due to the resolution limit. Subsequently, the accuracy of the statistical analysis decreases with decreasing blood vessel size (smaller blood vessels < 10 micrometers) above the circle of Willis).
Various over-the-counter medications and supplements were used by the subjects over the time period of this study; however, the BP changes caused by these medications should be minimal. Nonetheless, larger sample sizes are required to establish definitive relationship with progression to HBP. Despite these limitations, our method is relevant to understanding brain pathology relevant to hypertension whether such pathology precedes or follows the establishment of clinical hypertension.
Although it is hard to ask participants in a study that lasts for more than two years to visit the lab periodically to get their blood pressure measured, we suggest that taking blood pressure readings more frequently (maybe every three months) could provide more accurate observation of the MAP while excluding its possible temporal fluctuations, which could enhance the accuracy of this study.
Alterations in the cerebral vascular tree could be non-invasively detected by the analysis of MRA imaging data. Cerebrovascular changes are correlated to MAP below the circle of Willis. The improved segmentation algorithm coupled with the CDF and PDF estimations could indicate cerebral vascular alterations, which could assist clinicians to propose appropriate medical plans of hypertension.
Graph: Figure 1 Framework of the data analysis for quantifying cerebrovascular changes from MRA imaging data.
Graph: Figure 2 (a) Original MRA image slice of sample patient at day 0. (b) Manually segmented ground truth (GT) image from image in (a).
Graph: Figure 3 Illustration of segmentation accuracy and errors of the proposed automatic segmentation (C) by comparing to the ground truth (G) [[
Graph: Figure 4 Example of segmentation algorithm output. (a) Sample image slices of a patient at day 0. (b) Sample 3D reconstruction of the segmented cerebrovascular system using a growing tree model.
Graph: Figure 5 Sample patient CDFs demonstrating the temporal changes from day 0 to 700. The graphs indicate the probability that blood vessels may be of a certain diameter or less. (A–D) represents 4 different patients from our dateset.
MAP: Figure 6 Applying a 3D growing algorithm to the volume of binary segmented images allows for visualization of the automatically segmented MRA data. Day 0 of patient C (top left). Day 700 of patient C (top right). Day 0 of patient D (bottom left). Day 700 of patient D (bottom right). The correlation between the CDF and MAP was not found in these patients. Some apparent differences between vascular constructions in areas below the circle of Willis are highlighted.
Table 1 Sensitivity and specificity values for automatically segmented images at day 0, day 700, and cumulative.
Time Sensitivity Specificity Day 0 0.997 ± 0.006 0.9998 ± 0.0001 Day 700 0.996 ± 0.008 0.9998 ± 0.0001 Cumulative 0.997 ± 0.008 0.9998 ± 0.0001
Table 2 Statistical evaluation mixed effects linear model, where p-values < 0.05 were considered statistically significant. Diameter denotes size of vasculature in segmentation images. Age, gender, and timepoints are clinically acquired data.
Age 3.2 μm/y 0.356 0.551 Gender F > M by 12.8 μm 0.026 0.872 Mean Arterial Pressure −5.3 μm/mmHg 11.63 0.0007 Age −16.5 μm/y 12.29 0.0005 Gender F > M by 16.0 μm 0.199 0.655 Mean Arterial Pressure 1.6 μm/mmHg 0.402 0.525
Table 3 BP measurements of patients A, B, C, and D.
Patient Day 0 Day 700 Systolic BP Diastolic BP MAP Systolic BP Diastolic BP MAP A 120 80.5 93.7 103.5 66.5 78.8 B 130.5 83 98.8 143.5 94 110.5 C 118 80.5 93 105.3 69 81.1 D 114 84.5 94.3 120 88 98.7
Conceptualization, F.T., H.K., A.M., A.S. and A.E.-B.; data curation, F.T.; formal analysis, F.T. and A.E.-B.; investigation, F.T., Y.G. and A.E.-B.; methodology, H.K., A.M., A.S. and A.E.-B.; project administration, A.E.-B.; supervision, S.E.-M. and A.E.-B.; validation, F.T. and S.E.-M.; writing—original draft, F.T., H.K., Y.G., A.M. and A.S.; writing—review & editing, F.T., H.K., Y.G., S.E.-M. and A.E.-B. All authors have read and agreed to the published version of the manuscript.
This research received no external funding.
Data used in this study were collected and approved by the Institutional Review Board of Pittsburgh University.
Informed consent was obtained from all subjects involved in the study.
The data that support the findings of this study are available from the corresponding author upon reasonable request.
The authors declare no conflict of interest.
- I. Initialization sequentially using EM Algorithm
Consider F being the marginal distribution of gray level, EM algorithm [[
A mixture P
Subordinate components of the LCDG alternatingly approximate the deviations between F and P
Separating and scaling up the positive and the negative deviations to get the two probability distributions, D
Iteratively finding subordinate mixtures of positive and negative DGs using the same EM algorithm. These mixtures should best approximate D
Scaling down the positive and negative subordinate mixtures and adding them to the dominant mixture, which gives the initial LCDG model whose size is C = C
The initial LCDG has K dominant weights w
- II. Refining LCDGs using Modified EM Algorithm
The initial LCDG was refined by estimating the local maximum of the log-likelihood in (
Consider
taking into consideration that the following constraints apply:
Two main steps are repeated iteratively until the log-likelihood is maximized, the E-step
E-step
M-step
The described process was shown to be converging to a local log-likelihood maximum, using similar considerations as in [[
Considering unit factor constraints in (A2), the log-likelihood in (
Using (
where c can be either p or n. The modified EM-algorithm is true till the weights w become strictly positive. The iterations must be ended if the log-likelihood of (
Associating the subordinate DGs with the dominant terms, the final mixed LCDG-model
By Fatma Taher; Heba Kandil; Yitzhak Gebru; Ali Mahmoud; Ahmed Shalaby; Shady El-Mashad; Ayman El-Baz; Anton Civit and Sos Agaian
Reported by Author; Author; Author; Author; Author; Author; Author; Author; Author