Automatic Quantification of Enlarged Perivascular Space in Patients With Traumatic Brain Injury Using Super-Resolution of T2-Weighted Images

Abstract

The perivascular space (PVS) is important to brain waste clearance and brain metabolic homeostasis. Enlarged PVS (ePVS) becomes visible on magnetic resonance imaging (MRI) and is best appreciated on T2-weighted (T2w) images. However, quantification of ePVS is challenging because standard-of-care T1-weighted (T1w) and T2w images are often obtained via two-dimensional (2D) acquisition, whereas accurate quantification of ePVS normally requires high-resolution volumetric three-dimensional (3D) T1w and T2w images. The purpose of this study was to investigate the use of a deep-learning-based super-resolution (SR) technique to improve ePVS quantification from 2D T2w images for application in patients with traumatic brain injury (TBI). We prospectively recruited 26 volunteers (age: 31 ± 12 years, 12 male/14 female) where both 2D T2w and 3D T2w images were acquired along with 3D T1w images to validate the ePVS quantification using SR T2w images. We then applied the SR method to retrospectively acquired 2D T2w images in 41 patients with chronic TBI (age: 41 ± 16 years, 32 male/9 female). ePVS volumes were automatically quantified within the whole-brain white matter and major brain lobes (temporal, parietal, frontal, occipital) in all subjects. Pittsburgh Sleep Quality Index (PSQI) scores were obtained on all patients with TBI. Compared with the silver standard (3D T2w), in the validation study, the SR T2w provided similar whole-brain white matter ePVS volume (r = 0.98, p < 0.0001), and similar age-related ePVS burden increase (r = 0.80, p < 0.0001). In the patient study, patients with TBI with poor sleep showed a higher age-related ePVS burden increase than those with good sleep. Sleep status is a significant interaction factor in the whole brain (p = 0.047) and the frontal lobe (p = 0.027). We demonstrate that images produced by SR of 2D T2w images can be automatically analyzed to produce results comparable to those obtained by 3D T2 volumes. Reliable age-related ePVS burden across the whole-brain white matter was observed in all subjects. Poor sleep, affecting the glymphatic function, may contribute to the accelerated increase of ePVS burden following TBI.

Introduction

The perivascular spaces (PVS) are pial-lined passageways around vascular structures in the brain.¹ They are a neuroanatomical route for fluid flow that is crucial for brain waste clearance and brain metabolic homeostasis. The PVS are considered part of a “glymphatic” system, where cerebrospinal fluid (CSF) and brain interstitial fluid exchange, clearing interstitial solutes from the brain into the bloodstream. PVS typically become visible on magnetic resonance imaging (MRI) only when enlarged and are visualized as hyperintense foci on T2-weighted (T2w) images. It has been hypothesized that the increased visibility of PVS is reflective of impaired glymphatic function.^1,2 Enlargement of PVS (ePVS) has been reported with many conditions associated with impaired glymphatic function, such as normal aging,^3,4 small vessel disease,⁵ dementia, and Alzheimer's disease,^6,7 and multiple sclerosis (MS).⁸

The glymphatic system is disengaged during wakefulness and most active during sleep. This is accomplished by the expansion of the brain's extracellular space during natural sleep permitting CSF infiltration along the PVS.⁹ MRI-derived ePVS burden has been shown to correlate negatively with objective measures of sleep quality and may serve as a marker for nocturnal metabolite clearance.¹⁰ Given that 30–70% of patients with mild TBI report sleep problems^11,12, increased ePVS burden on MRI has been recently reported in patients with TBI as well, specifically linked to poor sleep and persistent post-concussive symptom severity.^13,14

However, quantification of ePVS is challenging, as manual delineation is too time consuming for clinical use. ePVS is typically rated by severity categories based on the approximate number of ePVS foci identified as hyperintensity spots on T2w MRI images¹⁵ in an anatomically defined region. However, such qualitative scores can easily have floor and ceiling effects, affecting their reliability.¹ Recent technological advancements have allowed the development of automatic and semiautomatic ePVS segmentation and quantification methods, mostly relying on intensity thresholding and image processing filters to extract vessel-like structures within the brain.^16,17 More recently, Sepehrband and colleagues proposed a fully automatic method to enhance the ePVS visibility through image processing, followed by vessel-like structure segmentation with the Frangi filter.¹⁸ The performance of this automatically quantified ePVS volume was validated in comparison with those manually counted by an expert.

Due to the small size of ePVS, most automatic segmentation techniques require three-dimensional (3D) high-resolution T1w and T2w images for the best quantification results. Whereas 3D T1w acquisitions of close to 1-mm isotropic resolution are frequently available in routine clinical exams, T2w images are more often acquired with two-dimensional (2D) thick slices (e.g., 5–6 mm), sometimes even with slice gaps to reduce image acquisition time. This lower-resolution T2w acquisition poses limitations to implementing the automatic ePVS quantification techniques on clinically acquired data sets or large retrospective data sets.

Super-resolution (SR) techniques have been developed to produce 3D volumetric data from stacks of 2D images. Traditional approaches include combining multiple stacks of MR images from different orientations to form a high-resolution 3D volume.^19,20 Recent developments in machine learning and deep learning have led to self-SR techniques that can produce high-resolution images from a single stack of 2D slices through learning-based algorithms.^21
–23 With the latest advances in deep-learning innovations for SR, the radiology community has shown increasing interest in exploring the clinical applications of SR techniques to enhance 3D volumetric imaging.²⁴

Several studies have reported promising findings regarding the clinical utilization of SR in the context of medical imaging. Steeden and associates employed a 3D residual U-Net for the SR reconstruction of rapidly acquired low-resolution 3D whole-heart cardiac MR images.²⁵ The model was trained with synthetic data from 500 high-resolution whole-heart data sets. The resulting SR high-resolution whole-heart images demonstrated a significant improvement in diagnostic accuracy and confidence compared with the low-resolution images. Importantly, the performance of the SR images was comparable to that of reference high-resolution data, which required three times longer to acquire. Zhou and colleagues conducted an evaluation of an MRI-based brain-tumor SR generative adversarial network (MRBT-SR-GAN) model on T2 fluid-attenuated inversion recovery (FLAIR) images.²⁶ The study demonstrated that the SR T2-FLAIR images effectively recovered high-resolution MRI details, leading to improved brain-tumor segmentation results. This advancement holds the potential for facilitating early detection and accurate evaluation of brain tumor recurrence and prognosis. Further, Hou and co-workers employed a deep-transfer-learning network super-resolve slice resolution from 3 mm to 0.825 mm on T2w MR images.²⁷ The authors demonstrated that the radiomics model trained with SR T2w images outperformed the original model based on conventional thick-slice T2w images in predicting tumor T-staging and identifying neoadjuvant chemoradiotherapy candidates among patients with rectal cancer.

Typical deep-learning-based SR methods require a large amount of external training data, and the SR results depend heavily on similarities between the training data and subject data. Our group recently proposed a self-supervised SR technique, SMORE (Synthetic Multi-Orientation Resolution Enhancement), a self-supervised anti-aliasing and SR algorithm for MRI using deep learning.^28,29 SMORE is a self-SR algorithm that uses the high-resolution in-plane slices to train a convolutional neural network (CNN)-based SR network and then applies it in the through-plane directions. The training data for SMORES are created by blurring the in-plane patches to match the through-plane resolution, with the full width at half max (FWHM) of the blurring filter designed based on the estimated slice selective profile by ESPRESO (Estimating the Slice Profile for Resolution Enhancement of a Single image Only), a method based on generative adversarial networks (GANs).³⁰ Because it does not use external training data and does not need to know the slice profile, the combined SMORE/ESPRESO method can be used to super-resolve virtually any 2D-acquired MRI.

In this study, we hypothesized that by application of SMORE to 2D T2w images, enlarged PVS could be characterized more precisely and that these results could be used to provide insights into the role of the glymphatic system during TBI recovery. Further studies might then use this technique on larger studies using clinical imaging data, leading to new targets for therapeutics. Therefore, we investigated the use of SMORE to super-resolve 2D T2w images to aid in ePVS quantification. We first conducted a validation study where 3D T1w, 3D T2w, and 2D T2w images were acquired from 30 healthy volunteers. The ePVS quantified using the SR 2D T2w images (with and without SMORE processing) were compared with 3D T2w images as the silver standard. We then applied the SMORE processing and ePVS quantification on a previously acquired data set on 41 patients with TBI with only 2D T2w images. As the glymphatic system is believed to be most active during sleep for waste clearance, and recent studies have indicated sleep disruption may be the cause for increased ePVS burden^13,14 following TBI, we also hypothesized that poor sleep quality in patients with TBI will be associated with increased ePVS burden in the brain from SMORE-derived volumetric data.

Methods

Participants

Participants in the study were recruited into two separate arms. The first arm, which we term the “validation study,” was designed to validate the feasibility of using the 2D T2w images to estimate ePVS. For this validation study, we recruited 26 healthy volunteers (age: 31 ± 12 years, range: 18–60 years, 12 male/14 female) aged 12–60 years with no history of neurological diseases such as seizures, stroke, TBI or any other condition requiring medical attention within the last 5 years. Volunteers were recruited through online advertisements in the University of Maryland Baltimore newsletter and flyers posted on the university campus. They were screened over the phone through designed screening checklists.

The second arm of the study, which we term the “patient study,” included 41 patients with TBI with a comparable age range at 6–18 months post-injury from the MagNeTS study (Magnetic Resonance Imaging of NeuroTrauma Study).³¹ Participant information is provided in Table 1. Patient sleep quality was also available from the study and was measured using the Pittsburgh Sleep Quality Index (PSQI).³² The PSQI contains 19 self-rated items that measure seven components of sleep quality (i.e., subjective sleep quality, sleep latency, sleep duration, habitual sleep efficiency, sleep disturbances, use of sleep medications, and daytime dysfunction), with a total score ranging from 0 to 21. A global PSQI score >5 was considered indicative of poor sleep.¹⁴

Table 1.

Demographic and Clinical Information of All Patients

	All	Good sleep	Poor sleep	P-value (good vs. poor sleep)
N	41	17	24
Age (years)	41 ± 16	44 ± 16	39 ± 17	0.25
Sex (M/F)	32/9	13/4	19/5	0.84
Race (B, W, O)	W = 31/B = 9/O = 1	W = 12/B = 5	W = 19/B = 4/O = 1	0.78
Education (years)	14 ± 3	13 ± 2	14 ± 2	0.11
GCS score	11.7 ± 4.5	11.7 ± 4.9	11.8 ± 4.3	0.97
Days post-injury	327 ± 171	353 ± 197	306 ± 151	0.47
Positive CT	19 (46%)	9 (52%)	10 (42%)	0.45
Positive MRI	17 (41%)	9 (52%)	8 (33%)	0.2
PSQI (0–21)	7.0 ± 4.3	3.6 ± 1.5	9.5 ± 4.0	< 0.00001^*

Patient sleep quality was measured by the PSQI, where patients with a global PSQI score >5 were considered as with poor sleep. Bold text indicate p < 0.05.

B, Black or African American; CT, computed tomography; GSC, Glasgow Coma Scale; MRI, magnetic resonance imaging; O, unknown; PSQI, Pittsburgh Sleep Quality Index; W, white.

The Institutional Review Board of the university approved this study. All participants signed the written informed consent with HIPAA compliance.

Data acquisition

Validation study

To test whether transforming the 2D T2w images using SMORE would provide equivalent ePVS estimation as the natively acquired 3D T2w images, we obtained 3D T1w, 3D T2w, and 2D T2w acquisitions during the same imaging session on a 3T Siemens Prisma scanner with a 64-channel head and neck coil. Both the 3D T1 and T2w images were acquired with 1 mm² isotropic resolution, 256 mm field of view (FOV), and 176 sagittal sections. Other imaging parameters for 3D T1w images were echo time (TE)/repetition time (TR)/inversion time (TI) = 3.37 msec/4000 msec/1400 msec, flip angle = 6 degrees, BW = 200 Hz/Px. For 3D T2w images, the acquisition parameters were TE/TR = 349 msec/3200 msec, turbo factor 280, and BW = 781 Hz/Px. The 2D T2w sequence was taken from a standard clinical exam protocol with 26 axial slices (5 mm thickness with 1 mm gap) at 230 mm FOV, TE/TR = 85 msec/5820 msec, in-plane resolution 0.4 mm × 0.4 mm, turbo factor 15, and BW = 195 Hz/Px.

Patient study

All data were obtained on a Siemens Tim Trio 3T MRI scanner with a 32-channel head coil. The 3D T1w image was acquired with 1 mm³ isotropic resolution, 256 mm FOV, and 196 sagittal sections, TE/TR/TI = 2.91 msec/2300 msec/900 msec, flip angle = 9 degrees, BW = 240 Hz/Px. There are no 3D T2w data available for this data set. The 2D T2w image was acquired with 26 axial slices (5 mm thickness with 1 mm gap) at 230 mm FOV, TE/TR = 85 msec/5690 msec, in-plane resolution 0.4 mm × 0.4 mm, turbo factor 15, BW = 195 Hz/Px.

Image processing and statistical analysis

ePVS segmentation

The 3D T1w and T2w images were processed using the Human Connectome Project (HCP) pipeline,³³ which includes registration between the T1w and T2w images, and sub-cortical segmentation with the FreeSurfer software (version v6.0.1). T1w and T2w volumes were further intensity normalized based on white-matter signal intensity using the histogram normalization module available within the Quantitative Imaging Toolkit (QIT).³⁴ ePVS quantification was performed as described by Sepehrband and colleagues,¹⁸ with software provided by Dr. Sepehrband. An enhanced perivascular contrast (EPC) image was first calculated by dividing the non-local adaptive filtered images (T1w/T2w). ePVS appears bright on T2w images and dark on T1w images. Through the image division, the EPC image shows intensified dark regions, with better contrast and visibility from surrounding white matter regions. A Frangi filter was then applied to the EPC images to generate a vessel-likeliness map, followed by thresholding to generate the final ePVS masks. The threshold “h” was set at 0.0000003 as recommended¹⁸ and was also found to provide a good balance between sensitivity in detecting ePVS and limiting the false detection rate. ePVS volumes were then calculated within the whole-brain white matter, as well as in regional brain lobes (frontal, parietal, temporal, and occipital) based on the FreeSurfer segmented masks (following the recon-all command).

ePVS can also be quantified with the 3D T1w images only when no T2w images are available, where the filtered T1w images were used in place of the EPC images to quantify hypointensity “vessel-like” regions on the T1w images.³⁵ We performed ePVS quantification with 3D T1w images only as a comparison for cases in the absence of T2w acquisition. The threshold for the vessel likeliness map for T1w only estimation was chosen as 0.000002, following the same principle as outlined above.³⁵

SR volumetric transformation of the 2D T2w images

Images obtained as T2w acquisition were first processed using the SMORE pipeline²⁹ to generate SR T2w volume images. The SR T2w images thus obtained at 1 mm thick were used for the ePVS quantification pipeline as outlined above.

As a comparison to the SR T2w, a 3D T2w volume was also created by registering the 2D T2w images to the 3D T1w volume using rigid transformation (6 degrees of freedom) first and then sinc-interpolation (Reg T2w) using FLIRT (FMRIB's Linear Image Registration Tool, FMRIB Software Library v6.0).³⁶ This sinc-interpolated image volume served as a comparison to the SR image volume generated from SMORE. ePVS segmentation was performed with a vessel likeliness threshold of 0.0000002 for both SR T2w and Reg T2w images, which is slightly lower than the threshold used for standard 3D T2w images in order to accommodate for the slight degradation in clarity of ePVS boundaries.

Validation study

Ideally, the gold standard of ePVS volumes should be the “true” volume of ePVS within the brain, which is not easily obtainable. We, therefore, used a silver standard³⁷ in the validation study defined as the ePVS volume calculated from the pair of the acquired 3D T1w and 3D T2w images. This allowed the comparison of all estimated ePVS volumes from various image types to the silver standard. Pearson correlation and intraclass correlation (ICC) were assessed between the whole-brain white matter ePVS volume from SR T2w, Reg T2w, and 3DT1w with the silver standard from each subject. ICC estimates and their 95% confidence intervals were calculated based on a single-rating (k = 1), consistency, and two-way mixed-effect model. We further explored the relationship of ePVS volume with age and ePVS distribution within different brain lobes. To normalize the brain size differences, the whole-brain ePVS burden was calculated as: ePVS burden = 100% × ePVS volume/whole-brain white matter volume. The relationship between the ePVS burden and age was investigated through the Pearson correlation of whole-brain ePVS burden with age. For regional ePVS burden, the ePVS volume within each brain lobe was normalized by the corresponding brain-lobe white matter volume.

Patient study

In the patient study, 2D T2w images were first super-resolved, becoming the SR T2w volume, and then combined with the 3D T1w images for ePVS quantification. ePVS burden was calculated in the same manner as described above. Multi-variate linear regression was used to assess the relationship between the ePVS burden with age and sleep as outcomes (i.e., PSQI global score). Statistical analysis was performed in R (R package version 1.0.44). Significance was defined as p < 0.05.

Results

Validation of ePVS quantification using SR T2w data

ePVS visualization and segmentation

Figure 1 shows the coronal and axial views from a 3D T2w image, a sinc-interpolated 2D T2w image (Reg T2w), and a SMORE SR image (SR 3DT2w), all from the same subject following rigid registration. Qualitatively, the SR 3DT2w images were far superior to the sinc-interpolated 2D T2w images, showing both improved structural clarity in-plane and much improved resolution in the slice direction.

FIG. 1.

Coronal and axial views of the 2D T2w with SMORE SR T2w technique and rigid registration with sinc interpolation (Reg T2w), as compared with the 3D T2w images. The entire brain was acquired in the 3D acquisitions, with the sides cut off for display purposes. 2D, two-dimensional; 3D, three-dimensional; SMORE, Synthetic Multi-Orientation Resolution Enhancement; SR, super-resolution; T2w, T2-weighted.

Figure 2 shows the EPC images along with segmented ePVS masks in red obtained for different T1w and T2w image types: silver standard (3D T2w + 3D T1w), SR T2w (+ 3D T1w), Reg T2w (+ 3D T1w), and 3D T1w only. The EPC images are of T1/T2 contrast and therefore highlight the low T1w contrast, made further darker by dividing the high T2w contrast into the ePVS regions. Note that EPC enhances the contrast of ePVS regions much better with T2w images (silver standard, SR T2w, Reg T2w) compared with the T1w only image. SR T2w also provided similar visualization of ePVS as the silver standard images with the segmented ePVS mask more in concordance with the silver standard.

FIG 2.

The EPC images and the automatic segmented ePVS masks on a representative subject with different images: silver standard (3D T1w+3D T2w), SR T2w (3D T1w+SR T2w), Reg T2w (3D T1w+Reg T2w), and T1w only (3D T1w). In the T1 only image, the EPC represents the 3D T1w image only. Ellipsoid highlighted the frontal lobe region with different ePVS visualization and segmentation results. 2D, two-dimensional; 3D, three-dimensional; EPC, enhanced perivascular contrast; ePVS, enlarged perivascular space; Reg T2w, sinc-interpolated 2D T2w; SR, super-resolution; T1w, T1-weighted; T2w, T2-weighted.

ePVS quantification

Figure 3 shows the calculated total ePVS volume within the whole-brain white matter for SR T2w, Reg T2w, and T1w only, as compared with the silver standard obtained from 3D T2w images. All methods showed a high correlation (p < 0.0001) with the silver standard ePVS volume, with SR T2w and Reg T2w performing better than T1 only. However, it should be noted that the estimated ePVS volume was lower with all the methods as compared with the silver standard, with the T1w only method being the worst. Table 2 shows the ICC of the quantified ePVS volume using different methods as compared with the silver standard. SR T2w has the highest ICC, 0.799, followed by Reg T2w (ICC = 0.699). The T1 only method had the lowest ICC, 0.567.

FIG. 3.

Whole-brain WM ePVS volume estimated using SR T2w, Reg T2w, and T1 only, as compared with the silver standard with 3D T2. Also shown are the Pearson correlation coefficient r and significance level p. 3D, three-dimensional; ePVS, enlarged perivascular space; Reg T2w, sinc-interpolated 2D T2w; SR, super-resolution; T2w, T2-weighted; WM, white matter.

Table 2.

Intraclass ICC Statistics

PVS method	ICC	95% CI	95% CI	F score	P-value
PVS method	ICC	Lower bound	Upper bound	F score	P-value
SR T2w	0.799	0.601	0.904	8.94	<0.0001
T1w only	0.567	0.238	0.78	3.62	0.001
Reg T2w	0.699	0.434	0.853	5.65	<0.0001

Table shows ICC statistics of the consistency of whole-brain white matter ePVS quantification using SR T2w, Reg T2w, and T1w only, as compared with the silver standard with 3D T2.

3D, three-dimensional; CI, confidence interval; ePVS, enlarged perivascular space; ICC, intraclass correlation coefficient; PVS, perivascular space; Reg T2w, sinc-interpolated 2D T2w; SR, super-resolution; T1w, T1-weighted; T2w, T2-weighted.

ePVS burden in the brain

Age-related whole-brain ePVS burden using various image types to arrive at the ePVS is shown in Figure 4. As expected, the whole-brain ePVS burden has a strong positive correlation with age. The SR T2 showed an age-related ePVS burden pattern most comparable to the silver standard with r = 0.80, p < 0.0001. Due to the general underestimation of the ePVS volume with any of the image types compared with the silver standard, the estimated linear coefficient of yearly ePVS burden increase among normal subjects is also lower with SR T2w (0.0057%), Reg T2w (0.0034%), and T1w only (0.0027%) methods, with SR T2w being the closest to the silver standard (0.0084%)

FIG. 4.

Age-related whole-brain WM ePVS burden among normal volunteers using various methods. ePVS burden was calculated as the whole-brain ePVS volume normalized by the whole-brain WM volume. Also shown is the estimated linear equation with x = age (years) and y = ePVS burden (%), along with Pearson correlation coefficient r and significance level p. 2D, two-dimensional; ePVS, enlarged perivascular space; Reg T2w, sinc-interpolated 2D T2w; SR, super-resolution; T2w, T2-weighted; WM, white matter.

Figure 5 shows a boxplot of regional ePVS burden in different brain lobes (frontal, occipital, parietal, and temporal lobes) as compared with the whole-brain ePVS burden. From the silver standard, it appears that the ePVS burden is the largest within the parietal lobe, followed by the frontal lobe. Both the occipital lobe and temporal lobe had similar ePVS burden. All three image types (SR T2w, T1 only, and Reg T2w) had lower whole-brain and lobe ePVS burden compared with the silver standard. However, the ePVS burden from the SR T2w was closest to the silver standard. In addition, the SR T2w also shows a similar trend of regional ePVS burden distribution as the silver standard, whereas the T1 only shows artificially high ePVS in the occipital lobe. The regional ePVS burden for Reg T2w appeared similar to that of SR T2w and the silver standard but had stronger outlier cases in the parietal lobe.

FIG. 5.

Global and regional ePVS burden in different brain lobes (frontal, occipital, parietal, and temporal lobes) with various methods. Regional ePVS burden was calculated as the ePVS volume within the lobe normalized by the WM volume within the corresponding region. ePVS, enlarged perivascular space; Reg T2w, sinc-interpolated 2D T2w; SR, super-resolution; T2w, T2-weighted; WM, white matter.

Patient study

The EPC images of patients with TBI were computed using 3D T1w images and SR T2w images, which in turn were derived from the 2D T2w images because this cohort lacked 3D T2w images. Figure 6 shows enhanced ePVS visualizations and segmentations using the SR T2 and EPC images as compared with T1w images only. Figure 7 shows a boxplot of regional ePVS burden in TBI patients in comparison to the volunteers' silver standard data from the validation study. The ePVS burden in all regions is higher among the patients with TBI compared with that of the volunteers. In addition, the relative ePVS burden distribution across lobes appears to be comparable with the volunteers, where the occipital lobe still shows the least ePVS burden, and the frontal and parietal lobes have higher ePVS burden. Table 3 shows the ePVS volumes and burdens within the whole-brain white matter and different brain lobes.

FIG. 6.

ePVS segmentation results on a representative patient with T1w images only (A) and with the EPC images composed from the SR T2w and T1w images. The automatically segmented ePVS mask is shown. The ellipsoid and arrow highlight regions with enhanced ePVS visualization and segmentation using the SR T2w images. EPC, enhanced perivascular contrast; ePVS, enlarged perivascular space; SR, super-resolution; T1w, T1-weighted; T2w, T2-weighted.

FIG. 7.

Global and regional ePVS burden in patients with TBI in comparison to healthy volunteers in different brain lobes (frontal, occipital, parietal, and temporal lobes). Regional ePVS burden was estimated as the ePVS volume within the lobe normalized by the WM volume within the corresponding region. ePVS, enlarged perivascular space; TBI, traumatic brain injury; WM, white matter.

Table 3.

ePVS Volume and Burden Within the Whole-Brain White Matter and in Different Brain Lobes

	Whole brain		Frontal lobe		Occipital lobe		Parietal lobe		Temporal lobe
	Volume (mm³)	Burden (%)	Volume (mm³)	Burden (%)	Volume (mm³)	Burden (%)	Volume (mm³)	Burden (%)	Volume (mm³)	Burden (%)
Mean	1657	0.39	601	0.40	57	0.11	412	0.34	185	0.25
SD	1104	0.28	467	0.33	60	0.14	291	0.26	139	0.21
Minimum	449	0.10	91	0.06	7	0.01	70	0.06	45	0.05
Maximum	5495	1.48	2366	1.68	365	0.87	1369	1.29	672	1.05

ePVS volumes are presented as mm³. ePVS burdens are shown as % ePVS volume over the whole brain or regional volume.

ePVS, enlarged perivascular space; SD, standard deviation.

Figure 8 shows the age-related whole-brain white matter ePVS burden in patients with TBI both with good and poor sleep. Although age is a significant factor in the ePVS burden in both groups, the ePVS burden was higher among patients with poor sleep. A multiple linear regression model shows sleep status is a significant factor for the age-related ePVS burden in the whole brain (p = 0.047) and the frontal lobe (p = 0.027).

FIG. 8.

Age-related whole-brain and frontal-lobe white matter ePVS burden in patients with TBI with good (PSQI ≤5, n = 18) and poor sleep (PSQI >5, n = 23). ePVS, enlarged perivascular space; PSQI, Pittsburgh Sleep Quality Index; TBI, traumatic brain injury.

Discussion

In this study, we examined (1) whether transforming 2D T2w imaging data into SR images using SMORE would be comparable with the native 3D T2w images when estimating ePVS, and (2) whether the automatic ePVS processing could differentiate ePVS burden among patients with mild TBI compared with control subjects, and whether the sleep quality among the patients with mild TBI is associated with changes in ePVS. Our results indicate that SR 2D T2w images can be used along with the 3D T1w images to improve automatic ePVS quantification in the brain. The innovative SMORE self-supervised SR technique works robustly to provide 3D volumetric information from stacks of thick 2D slices even when there are slice gaps. Validation results on 26 volunteers indicate that the SR T2w images provide results highly comparable to the natively acquired 3D T2w images in ePVS quantification in the whole-brain white matter. Further, results from patients with TBI indicate a strong age-related ePVS burden across the brain with sleep status (good vs. poor) being a significant modifying factor in this age-related ePVS burden.

Normally, the PVS cross-sectional diameter is less than 1.5 mm,³⁸ and is considered enlarged when the diameter exceeds 3 mm,³⁹ and becomes more visible on MRI. The PVS is visible on MR images because of its fluid content. PVS appears as lines when they are parallel to the imaging plane, and as dots when they run through the plane. Although the precise composition of the fluid within PVS is not completely known, the contrast from PVS is typical to that of CSF. Given the high T2 content in PVS, T2w imaging is typically favored for PVS visualization.⁴⁰ However, T2w images are more often acquired as 2D axial stacks in routine clinical exams, although more and more 3D T2w images are being acquired in some centers. Therefore, 3D T1w images that have higher spatial resolution and visualize CSF as a dark signal are often used in conjunction with 2D T2w images and FLAIR sequences to rule out other pathologies and in making a qualitative assessment of PVS burden.¹⁵

Our results from the volunteer validation aspect of this study agree with previous studies that show a significant age effect on the whole-brain ePVS burden.⁴¹ Enlarged PVS is generally thought to reflect a compensatory mechanism for the age-related decline in glymphatic function, therefore maintaining homeostasis in the brain.^42,43 Our results also indicate the ePVS burden to be the highest in the parietal lobe, followed by the frontal lobe. Although these results are from a limited data set of 26 volunteers, a similar trend was reported using the same segmentation methods in a much larger cohort of ∼1400 subjects in the Human Connectome Project.⁴⁴ Using a different segmentation method also based on 3D T1w and 3D T2w images, characterization of PVS burden within the white matter of 118 healthy adolescents also indicated a more common presence of PVS in the frontal and parietal lobes than in the temporal and occipital lobes.⁴⁵ Our results also show the same trend regarding the distribution of ePVS burden across the brain using both the silver standard and the SR T2w images, but such a trend was not evident when only T1w images were used in the estimation of ePVS. This highlights the importance of including the T2w images, even if it is a 2D acquisition, to obtain a more accurate ePVS estimation. Note a silver standard is used here, which is often considered in image segmentation literature when the gold standard manual segmentation labels are not available. The segmentation masks using the consensus algorithm are then used to validate other comparison methods.^37,46
–48

After successful verification that the SR 2D-T2w images provided similar results on ePVS quantification compared with those obtained from 3D-T2w images in a normal population, we performed ePVS quantification on a group of patients with chronic TBI from the legacy MagNeTS study, where only 2D T2w images were available. Once these images were super-resolved using SMORE, a similar age-related ePVS burden in the TBI population was observed in the healthy control population. Note the overall ePVS burden is higher in the patient population than in the volunteers. However, this may be confounded by the differences in imaging conditions associated with the two different scanner types (e.g., different scanner and signal-to-noise levels). More interestingly, we observed that sleep quality significantly altered the normal age-related ePVS burden, where poor sleep contributed to a stronger age-related ePVS burden increase. Enlarged PVS has been suspected as a hallmark of TBI as early as 2005.⁴⁹

Recent studies examining ePVS with other patient outcome measures have pointed toward a specific link between ePVS burden and sleep conditions following TBI. Opel and colleagues¹³ quantified ePVS burden by manual counting of ePVS on a single axial MRI slice containing the basal ganglia and centrum semiovale and reported a correlation between increased ePVS burden and decreased total sleep time in 38 participants (13 TBI and 25 no-TBI participants), and a stronger correlation with total sleep time, sleep efficiency, and percent wake time in the patients with TBI only. In a moderately sized cohort of 56 U.S. military veterans, Piantino and associates¹⁴ quantified white matter ePVS using an automatic method based on 3D T1w MR signal intensity and the expected morphological characterization (width, volume, linearity) of the ePVS region, along with 3D FLAIR images to rule out abnormal white matter regions.^50,51 Piantino and associates reported a higher ePVS burden with more military-related mTBIs, as well as an interaction between mTBI and poor sleep on PVS burden.¹⁴

The glymphatic function has been shown to be most active during deep sleep⁵² and therefore it has been hypothesized that sleep disturbances, being a common symptom following TBI, may contribute to disrupted glymphatic function. Indeed, sub-acute (∼1 month post-injury) sleep disturbances have been reported to be associated with poor long-term outcomes, such as poor concentration and memory problems, worse neurocognitive function, slower overall recovery, and lower satisfaction of life.⁵³ The results of this study and the preceding study collectively strengthen the case that sleep disturbances may lead to enhanced ePVS burden and therefore could serve as a reliable prognostic factor of long-term outcomes from TBI. It should be noted that there could be other causes for the PVS enlargement post-TBI (e.g., vascular dysregulation, etc.), which will require more extensive research.

The findings of this study must be viewed in the context of its limitations. The validation aspect of the study was limited to the whole-brain white matter regions and that was mainly due to a limitation of the automatic ePVS quantification algorithm. Sub-cortical regions such as basal ganglia often present as prominent regions with ePVS presence, but they are missing in the scope of this work. Future developments in quantification should also consider looking into ePVS quantification in these sub-cortical gray matter regions for a more comprehensive estimation of ePVS burden in the whole brain. In addition, the sensitivity to PVS contrast may be affected by the signal-to-noise ratio (SNR) of images, where noise may also appear as dark spots on the T1w images. It is possible that the overall higher ePVS burden in the TBI patient population may partly be attributed to the lower SNR in the images. However, we were able to observe a similar age-related ePVS burden and a similar brain lobe distribution of ePVS in the patients with TBI as in the silver standard data from healthy volunteers, which leads us to believe that SNR was less of an issue in the quantification of ePVS. Finally, our patient results were based on a small number of patients with chronic TBI, which limits the significance of our current findings. Findings from this preliminary patient study should be further explored with a larger data set of patients with TBI such as that from the TRACK-TBI (Transforming Research and Clinical Knowledge in Traumatic Brain Injury) study, in association with sleep quality and other outcome metrics.

Conclusions

In this study, we demonstrated that the innovative SMORE technique can super-resolve a stack of 2D T2w images to produce 3D T2 volumes (SR T2). The SR T2 volume provides comparable ePVS estimation as a natively acquired 3D T2 volume. Results in both healthy volunteers and patients with TBI indicate a strong age-related ePVS burden in various brain regions. Further, sleep quality at the chronic stage of TBI appears to be a significant modifying factor in age-related ePVS burden. Poor sleep, affecting the normal glymphatic function, may contribute to the accelerated increase of ePVS burden following TBI. Overall, SR T2 allows the extension of the automatic ePVS quantification to routine clinical exams where only 2D T2w images are acquired, and hence could be applied to previously acquired, large multi-center data with only 2D T2w images.

Transparency, Rigor, and Reproducibility Summary

The patient data was pre-registered, with more details provided in the article by Hou and co-workers.²⁷ Volunteer data were not formally registered because the study was performed as an imaging quality improvement project. The analysis plan was not formally pre-registered, but the team member with primary responsibility for the analysis (JZ) certifies that the analysis plan was pre-specified. For the patient study, the sample size of 41 patients was included based on the availability of imaging data at chronic time-points post-injury in the MagNeTS study. The observed effect size of the multiple linear regression with age-related changes in ePVS in the whole-brain white matter was large, with Cohen's f² = 0.5 (F = 36.5), and the age by sleep interaction was small to medium, with Cohen's f² = 0.08 (F = 4.23). For the validation study, 26 participants were included from whom images were collected and analyzed. Imaging data were labeled using subject IDs without any identifying information of the participants. Imaging data from the validation study were acquired from January 2021 to July 2021 using the same scanner (Siemens Prisma). Patient data from the MagNeTS study were acquired between March 2010 and March 2015 on two MRI scanners (both were Siemens Tim Trio).

All imaging data sets were analyzed at the same time. Complete imaging parameters are presented in the Methods section. Data pre-processing was performed using the HCP pipeline (https://github.com/Washington-University/HCPpipelines). ePVS quantification software may be available upon request from Dr. Sepehrband.¹⁸ The SMORE SR software is available for public download at Dr. Prince's website (https://iacl.ece.jhu.edu/index.php?title=Resources). The test-retest reliability of the primary clinical outcome measure has not been formally determined. The statistical tests used were based on the assumptions of normal distribution. As this is a report of image processing technique development, correction for multiple comparisons was not performed. Replication by the study group is ongoing. De-identified patient data from this study are available in the FITBIR public archive (https://fitbir.nih.gov/). The analytic code used to conduct the analyses presented in this study is not available in a public repository. It may be available by e-mailing the corresponding author. The authors agree to provide the full content of the manuscript through National Institutes of Health (NIH) public access.

Footnotes

Acknowledgments

The study was conducted at the University of Maryland School of Medicine Center for Innovative Biomedical Resources, Center for Translational Research in Imaging (CTRIM), Baltimore, Maryland. We thank Dr. Sepehrband for providing us with the PVS quantification tools.

Authors' Contributions

The authors contributed as follows. JZ: conceptualization, methodology, formal analysis, visualization, writing–original draft; PR: conceptualization, methodology, writing–review and editing; MS: software, validation, writing–review and editing; XL: investigation, data curation, writing–review and editing; SR: software, resources, project administration; RLNT: data curation, resources, writing–review and editing; CSR: conceptualization, methodology, writing–review and editing; NB: conceptualization, writing–review and editing, funding acquisition; JLP: conceptualization, writing–review and editing, funding acquisition; RPG: conceptualization, writing–review and editing, funding acquisition, supervision.

Funding Information

The study was funded by grants 5R01NS105503 and R03NS088014 from the National Institute of Neurological Disorders and Stroke, and by grants W81XWH-08-1-0725 and W81XWH-12-1-0098 from the U.S. Department of Defense.

Author Disclosure Statement

No competing financial interests exist.

References

Wardlaw

, Benveniste

, Nedergaard

, et al. Perivascular spaces in the brain: anatomy, physiology and pathology. Nat Rev Neurol, 2020; 16(3):137–153; doi: 10.1038/S41582-020-0312-Z

Ramirez

, Berezuk

, McNeely

, et al. Imaging the perivascular space as a potential biomarker of neurovascular and neurodegenerative diseases. Cell Mol Neurobiol, 2016; 36(2):289–299; doi: 10.1007/s10571-016-0343-6

Zhu

, Dufouil

, Mazoyer

, et al. Frequency and location of dilated Virchow-Robin spaces in elderly people: a population-based 3D MR imaging study. AJNR Am J Neuroradiol, 2011; 32(4):709–713; doi: 10.3174/AJNR.A2366

Francis

, Ballerini

, Wardlaw

. Perivascular spaces and their associations with risk factors, clinical disorders and neuroimaging features: a systematic review and meta-analysis. Int J Stroke, 2019; 14(4):359–371; doi: 10.1177/1747493019830321

Potter

, Doubal

, Jackson

, et al. Enlarged perivascular spaces and cerebral small vessel disease. Int J Stroke, 2015; 10(3):376–381; doi: 10.1111/IJS.12054

Zhu

, Dufouil

, Soumaré

, et al. High degree of dilated Virchow-Robin spaces on MRI is associated with increased risk of dementia. J Alzheimers Dis, 2010; 22(2):663–672; doi: 10.3233/JAD-2010-100378

Weller

, Hawkes

, Kalaria

, et al. White matter changes in dementia: role of impaired drainage of interstitial fluid. Brain Pathol, 2015; 25(1):63–78; doi: 10.1111/BPA.12218

Wuerfel

, Haertle

, Waiczies

, et al. Perivascular spaces—MRI marker of inflammatory activity in the brain?. Brain, 2008; 131(Pt 9):2332–2340; doi: 10.1093/BRAIN/AWN171

Reddy

, van der Werf

. The sleeping brain: harnessing the power of the glymphatic system through lifestyle choices. Brain Sci, 2020; 10(11):1–16; doi: 10.3390/BRAINSCI10110868

10.

Berezuk

, Ramirez

, Gao

, et al. Virchow-Robin spaces: correlations with polysomnography-derived sleep parameters. Sleep, 2015; 38(6):853; doi: 10.5665/SLEEP.4726

11.

Mathias

, Alvaro

. Prevalence of sleep disturbances, disorders, and problems following traumatic brain injury: a meta-analysis. Sleep Med, 2012; 13(7):898–905; doi: 10.1016/J.SLEEP.2012.04.006

12.

Castriotta

, Wilde

, Lai

, et al. Prevalence and consequences of sleep disorders in traumatic brain injury. J Clin Sleep Med, 2007; 3(4):349; doi: 10.5664/jcsm.26855

13.

Opel

, Christy

, Boespflug

, et al. Effects of traumatic brain injury on sleep and enlarged perivascular spaces. J Cereb Blood Flow Metab, 2019; 39(11):2258–2267; doi: 10.1177/0271678X18791632

14.

Piantino

, Schwartz

, Luther

, et al. Link between mild traumatic brain injury, poor sleep, and magnetic resonance imaging: visible perivascular spaces in veterans. J Neurotrauma, 2021; 38(17):2391–2399; doi: 10.1089/neu.2020.7447

15.

Potter

, Chappell

, Morris

, et al. Cerebral perivascular spaces visible on magnetic resonance imaging: development of a qualitative rating scale and its observer reliability. Cerebrovasc Dis, 2015; 39(3-4):224–231; doi: 10.1159/000375153

16.

Ballerini

, Lovreglio

, Valdés Hernández

MDC

, et al. Perivascular spaces segmentation in brain MRI using optimal 3D filtering. Sci Rep, 2018; 8(1); doi: 10.1038/S41598-018-19781-5

17.

Park

, Zong

, Gao

, Lin

, Shen

. Segmentation of perivascular spaces in 7T MR image using auto-context model with orientation-normalized features. Neuroimage, 2016; 134:223; doi: 10.1016/J.NEUROIMAGE.2016.03.076

18.

Sepehrband

, Barisano

, Sheikh-Bahaei

, et al. Image processing approaches to enhance perivascular space visibility and quantification using MRI. Sci Rep, 2019; 9(1); doi: 10.1038/s41598-019-48910-x

19.

Woo

, Bai

, Roy

, Murano

, Stone

, Prince

. Super-resolution reconstruction for tongue MR images. Proc SPIE Int Soc Opt Eng, 2012; 8314:83140C; doi: 10.1117/12.911445

20.

Plenge

, Poot

DHJ

, Bernsen

, et al. Super-resolution methods in MRI: can they improve the trade-off between resolution, signal-to-noise ratio, and acquisition time?. Magn Reson Med, 2012; 68(6):1983–1993; doi: 10.1002/MRM.24187

21.

, Song

, Guo

, et al. Super-resolution of brain MRI images using overcomplete dictionaries and nonlocal similarity. IEEE Access, 2019; 7:25897-25907; doi: 10.1109/ACCESS.2019.2900125

22.

Mahapatra

, Bozorgtabar

, Hewavitharanage

, et al. Image super resolution using generative adversarial networks and local saliency maps for retinal image analysis. Lecture Notes in Computer Science (LNCS), 2017; 382–390; doi: 10.1007/978-3-319-66179-7_44/FIGURES/4

23.

, He

, Wang

, et al. Super-resolution reconstruction of single anisotropic 3D MR images using residual convolutional neural network. Neurocomputing, 2020; 392:209–220; doi: 10.1016/J.NEUCOM.2018.10.102

24.

Rudie

, Gleason

, Barkovich

, et al. Clinical assessment of deep learning-based super-resolution for 3D volumetric brain MRI. Radiol Artif Intell, 2022; 4(2); doi: 10.1148/RYAI.210059

25.

Steeden

, Quail

, Gotschy

, et al. Rapid whole-heart CMR with single volume super-resolution. J Cardiovasc Magn Res, 2020; 22(1); doi: 10.1186/S12968-020-00651-X

26.

Zhou

, Ma

, Feng

, et al. Super-resolution of brain tumor MRI images based on deep learning. J Appl Clin Med Phys, 2022; 23(11); doi: 10.1002/ACM2.13758

27.

Hou

, Zhou

, Sun

. Deep-learning-based 3D super-resolution MRI radiomics model: superior predictive performance in preoperative T-staging of rectal cancer. Eur Radiol, 2023; 33(1):1–10; doi: 10.1007/S00330-022-08952-8/FIGURES/5

28.

Zhao

, Shao

, Carass

, et al. Applications of a deep learning method for anti-aliasing and super-resolution in MRI. Magn Reson Imaging, 2019; 64:132–141; doi: 10.1016/J.MRI.2019.05.038

29.

Zhao

, Dewey

, Pham

, et al. SMORE: a self-supervised anti-aliasing and super-resolution algorithm for MRI using deep learning. IEEE Trans Med Imaging, 2021; 40(3):805–817; doi: 10.1109/TMI.2020.3037187

30.

Han

, Carass

, Schar

Slice profile estimation from 2D MRI acquisition using generative adversarial networks. Proc Int Symp Biomed Imaging, 2021; 145–149; doi: 10.1109/ISBI48211.2021.9434137

31.

Sours

, Zhuo

, Roys

, et al. Disruptions in resting state functional connectivity and cerebral blood flow in mild traumatic brain injury patients. PLoS One, 2015; 10(8); doi: 10.1371/JOURNAL.PONE.0134019

32.

Buysse

, Reynolds

, Monk

, et al. The Pittsburgh Sleep Quality Index: a new instrument for psychiatric practice and research. Psychiatry Res, 1989; 28(2):193–213; doi: 10.1016/0165-1781(89)90047-4

33.

Glasser

, Sotiropoulos

, Wilson

, et al. The minimal preprocessing pipelines for the Human Connectome Project. Neuroimage, 2013; 80:105–124; doi: 10.1016/J.NEUROIMAGE.2013.04.127

34.

Cabeen

, Laidlaw

, Toga

Quantitative Imaging Toolkit: software for interactive 3D visualization, data exploration, and computational analysis of neuroimaging datasets. Proc Int Soc Magn Reson Imaging, 2018.

35.

Sepehrband

, Barisano

, Sheikh-Bahaei

, et al. Volumetric distribution of perivascular space in relation to mild cognitive impairment. Neurobiol Aging, 2021; 99:28–43; doi: 10.1016/j.neurobiolaging.2020.12.010

36.

Jenkinson

, Smith

. A global optimisation method for robust affine registration of brain images. Med Image Anal, 2001; 5(2):143–156; doi: 10.1016/S1361-8415(01)00036-6

37.

Lucena

, Souza

, Rittner

, et al. Convolutional neural networks for skull-stripping in brain MR imaging using silver standard masks. Artif Intell Med, 2019; 98:48–58; doi: 10.1016/J.ARTMED.2019.06.008

38.

Rudie

, Rauschecker

, Nabavizadeh

, et al. Neuroimaging of dilated perivascular spaces: from benign and pathologic causes to mimics. J Neuroimaging, 2018; 28(2):139–149; doi: 10.1111/JON.12493

39.

Ding

, Sigurðsson

, Jónsson

, et al. Large perivascular spaces visible on magnetic resonance imaging, cerebral small vessel disease progression, and risk of dementia: the Age, Gene/Environment Susceptibility–Reykjavik Study. JAMA Neurol, 2017; 74(9):1105; doi: 10.1001/JAMANEUROL.2017.1397

40.

Kwee

, Kwee

. Virchow-Robin spaces at MR imaging. Radiographics, 2007; 27(4):1071–1086; doi: 10.1148/RG.274065722

41.

Kim

, Shin

, Nam

, et al. MRI-visible dilated perivascular space in the brain by age: the Human Connectome Project. Radiology, 2023; 306(3):e213254; doi: 10.1148/radiol

42.

, Hu

, Li

, et al. Perivascular spaces, glymphatic system and MR. Front Neurol, 2022; 13:763; doi: 10.3389/FNEUR.2022.844938/BIBTEX

43.

Zhu

, Dufouil

, Mazoyer

, et al. Frequency and location of dilated Virchow-Robin spaces in elderly people: a population-based 3D MR imaging study. AJNR Am J Neuroradiol, 2011; 32(4):709–713; doi: 10.3174/AJNR.A2366

44.

Lynch

, Sepehrband

, Toga

, et al. Brain perivascular space imaging across the human lifespan. Neuroimage, 2023; 271:120009; doi: 10.1016/j.neuroimage.2023

45.

Piantino

, Boespflug

, Schwartz

, et al. Characterization of MR imaging-visible perivascular spaces in the white matter of healthy adolescents at 3T. Am J Neuroradiol, 2020; 41(11):2139–2145; doi: 10.3174/ajnr.A6789

46.

Pfefferbaum

, Sullivan

, Zahr

, et al. Multi-atlas thalamic nuclei segmentation on standard T1-weighed MRI with application to normal aging. Hum Brain Mapp, 2023; 44(2):612–628; doi: 10.1002/HBM.26088

47.

, Thomas

, Kasoff

, et al. Thalamus Optimized Multi Atlas Segmentation (THOMAS): fast, fully automated segmentation of thalamic nuclei from structural MRI. Neuroimage, 2019; 194:272–282; doi: 10.1016/J.NEUROIMAGE.2019.03.021

48.

Lucena

, Souza

, Rittner

, et al. Silver standard masks for data augmentation applied to deep-learning-based skull-stripping. Proc Int Symp Biomed Imaging, 2018; 1114–1117; doi: 10.1109/ISBI.2018.8363766

49.

Inglese

, Bomsztyk

, Gonen

, et al. dilated perivascular spaces: hallmarks of mild traumatic brain injury. AJNR Am J Neuroradiol, 2005; 26(4):719–724

50.

Schwartz

, Boespflug

, Lahna

, et al. Autoidentification of perivascular spaces in white matter using clinical field strength T1 and FLAIR MR imaging. Neuroimage, 2019:202; doi: 10.1016/J.NEUROIMAGE.2019.116126

51.

Boespflug

, Schwartz

, Lahna

, et al. MR Imaging–based Multimodal Autoidentification of Perivascular Spaces (mMAPS): automated morphologic segmentation of enlarged perivascular spaces at clinical field strength. Radiology, 2017; 286(2):632–642; doi: 10.1148/RADIOL.2017170205

52.

Nedergaard

, Goldman

. Glymphatic failure as a final common pathway to dementia. Science, 2020; 370(6512):50–56; doi: 10.1126/SCIENCE.ABB8739

53.

Tang

, Sours Rhodes

, Jiang

, et al. Association between sleep disturbances at subacute stage of mild traumatic brain injury and long-term outcomes. Neurotrauma Rep, 2022; 3(1):276–285; doi: 10.1089/neur.2022.0004