Application of the Ordered Logit Model to Optimising Frangi Filter Parameters for Segmentation of Perivascular Spaces

Systematic review and meta-analysis of automated methods for quantifying enlarged perivascular spaces in the brain

Research into magnetic resonance imaging (MRI)-visible perivascular spaces (PVS) has recently increased, as results from studies in different diseases and populations are cementing their association with sleep, disease phenotypes, and overall health indicators. With the establishment of worldwide consortia and the availability of large databases, computational methods that allow to automatically process all this wealth of information are becoming increasingly relevant. Several computational approaches have been proposed to assess PVS from MRI, and efforts have been made to summarise and appraise the most widely applied ones. We systematically reviewed and meta-analysed all publications available up to September 2023 describing the development, improvement, or application of computational PVS quantification methods from MRI. We analysed 67 approaches and 60 applications of their implementation, from 112 publications. The two most widely applied were the use of a morphological filter to enhance PVS-like structures, with Frangi being the choice preferred by most, and the use of a U-Net configuration with or without residual connections. Older adults or population studies comprising adults from 18 years old onwards were, overall, more frequent than studies using clinical samples. PVS were mainly assessed from T2-weighted MRI acquired in 1.5T and/or 3T scanners, although combinations using it with T1-weighted and FLAIR images were also abundant. Common associations researched included age, sex, hypertension, diabetes, white matter hyperintensities, sleep and cognition, with occupation-related, ethnicity, and genetic/hereditable traits being also explored. Despite promising improvements to overcome barriers such as noise and differentiation from other confounds, a need for joined efforts for a wider testing and increasing availability of the most promising methods is now paramount.

Deep Cascade of Convolutional Neural Networks for Quantification of Enlarged Perivascular Spaces in the Basal Ganglia in Magnetic Resonance Imaging

In this paper, we present a cascaded deep convolution neural network (CNN) for assessing enlarged perivascular space (ePVS) within the basal ganglia region using T2-weighted MRI. Enlarged perivascular spaces (ePVSs) are potential biomarkers for various neurodegenerative disorders, including dementia and Parkinson's disease. Accurate assessment of ePVS is crucial for early diagnosis and monitoring disease progression. Our approach first utilizes an ePVS enhancement CNN to improve ePVS visibility and then employs a quantification CNN to predict the number of ePVSs. The ePVS enhancement CNN selectively enhances the ePVS areas without the need for additional heuristic parameters, achieving a higher contrast-to-noise ratio (CNR) of 113.77 compared to Tophat, Clahe, and Laplacian-based enhancement algorithms. The subsequent ePVS quantification CNN was trained and validated using fourfold cross-validation on a dataset of 76 participants. The quantification CNN attained 88% accuracy at the image level and 94% accuracy at the subject level. These results demonstrate significant improvements over traditional algorithm-based methods, highlighting the robustness and reliability of our deep learning approach. The proposed cascaded deep CNN model not only enhances the visibility of ePVS but also provides accurate quantification, making it a promising tool for evaluating neurodegenerative disorders. This method offers a novel and significant advancement in the non-invasive assessment of ePVS, potentially aiding in early diagnosis and targeted treatment strategies.

Detectability and accuracy of computational measurements of in-silico and physical representations of enlarged perivascular spaces from magnetic resonance images

BACKGROUND

Magnetic Resonance Imaging (MRI) visible perivascular spaces (PVS) have been associated with age, decline in cognitive abilities, interrupted sleep, and markers of small vessel disease. But the limits of validity of their quantification have not been established.

NEW METHOD

We use a purpose-built digital reference object to construct an in-silico phantom for addressing this need, and validate it using a physical phantom. We use cylinders of different sizes as models for PVS. We also evaluate the influence of 'PVS' orientation, and different sets of parameters of the two vesselness filters that have been used for enhancing tubular structures, namely Frangi and RORPO filters, in the measurements' accuracy.

RESULTS

PVS measurements in MRI are only a proxy of their true dimensions, as the boundaries of their representation are consistently overestimated. The success in the use of the Frangi filter relies on a careful tuning of several parameters. Alpha= 0.5, beta= 0.5 and c= 500 yielded the best results. RORPO does not have these requirements and allows detecting smaller cylinders in their entirety more consistently in the absence of noise and confounding artefacts. The Frangi filter seems to be best suited for voxel sizes equal or larger than 0.4 mm-isotropic and cylinders larger than 1 mm diameter and 2 mm length. 'PVS' orientation did not affect measurements in data with isotropic voxels.

COMPARISON WITH EXISTENT METHODS

Does not apply.

CONCLUSIONS

The in-silico and physical phantoms presented are useful for establishing the validity of quantification methods of tubular small structures.

Influence of threshold selection and image sequence in in-vivo segmentation of enlarged perivascular spaces

BACKGROUND

Growing interest surrounds perivascular spaces (PVS) as a clinical biomarker of brain dysfunction given their association with cerebrovascular risk factors and disease. Neuroimaging techniques allowing quick and reliable quantification are being developed, but, in practice, they require optimisation as their limits of validity are usually unspecified.

NEW METHOD

We evaluate modifications and alternatives to a state-of-the-art (SOTA) PVS segmentation method that uses a vesselness filter to enhance PVS discrimination, followed by thresholding of its response, applied to brain magnetic resonance images (MRI) from patients with sporadic small vessel disease acquired at 3 T.

RESULTS

The method is robust against inter-observer differences in threshold selection, but separate thresholds for each region of interest (i.e., basal ganglia, centrum semiovale, and midbrain) are required. Noise needs to be assessed prior to selecting these thresholds, as effect of noise and imaging artefacts can be mitigated with a careful optimisation of these thresholds. PVS segmentation from T1-weighted images alone, misses small PVS, therefore, underestimates PVS count, may overestimate individual PVS volume especially in the basal ganglia, and is susceptible to the inclusion of calcified vessels and mineral deposits. Visual analyses indicated the incomplete and fragmented detection of long and thin PVS as the primary cause of errors, with the Frangi filter coping better than the Jerman filter.

COMPARISON WITH EXISTING METHODS

Limits of validity to a SOTA PVS segmentation method applied to 3 T MRI with confounding pathology are given.

CONCLUSIONS

Evidence presented reinforces the STRIVE-2 recommendation of using T2-weighted images for PVS assessment wherever possible. The Frangi filter is recommended for PVS segmentation from MRI, offering robust output against variations in threshold selection and pathology presentation.

Intensive Systolic Blood Pressure Treatment Remodels Brain Perivascular Spaces: A Secondary Analysis of SPRINT

Background

Brain perivascular spaces (PVS) are part of the glymphatic system and facilitate clearance of metabolic byproducts. Since enlarged PVS are associated with vascular health, we tested whether intensive systolic blood pressure (SBP) treatment affects PVS structure.

Methods

This is a secondary analysis of the Systolic PRessure INTervention (SPRINT) Trial MRI Substudy: a randomized trial of intensive SBP treatment to goal < 120 mm Hg vs. < 140 mm Hg. Participants had increased cardiovascular risk, pre-treatment SBP 130-180, and no clinical stroke, dementia, or diabetes. Brain MRIs acquired at baseline and follow-up were used to automatically segment PVS in the supratentorial white matter and basal ganglia using a Frangi filtering method. PVS volumes were quantified as a fraction of the total tissue volume. The effects of SBP treatment group and major antihypertensive classes on PVS volume fraction were separately tested using linear mixed-effects models while covarying for MRI site, age, sex, black race, baseline SBP, history of cardiovascular disease (CVD), chronic kidney disease, and white matter hyperintensities (WMH).

Results

For 610 participants with sufficient quality MRI at baseline (mean age 67±8, 40% female, 32% black), greater PVS volume fraction was associated with older age, male sex, non-Black race, concurrent CVD, WMH, and brain atrophy. For 381 participants with MRI at baseline and at follow-up (median = 3.9 years), intensive treatment was associated with decreased PVS volume fraction relative to standard treatment (interaction coefficient: -0.029 [-0.055 to -0.0029] p=0.029). Reduced PVS volume fraction was also associated with exposure to calcium channel blockers (CCB) and diuretics.

Conclusions

Intensive SBP lowering partially reverses PVS enlargement. The effects of CCB use suggests that improved vascular compliance may be partly responsible. Improved vascular health may facilitate glymphatic clearance. Clincaltrials.gov : NCT01206062.

A critical guide to the automated quantification of perivascular spaces in magnetic resonance imaging

The glymphatic system is responsible for waste clearance in the brain. It is comprised of perivascular spaces (PVS) that surround penetrating blood vessels. These spaces are filled with cerebrospinal fluid and interstitial fluid, and can be seen with magnetic resonance imaging. Various algorithms have been developed to automatically label these spaces in MRI. This has enabled volumetric and morphological analyses of PVS in healthy and disease cohorts. However, there remain inconsistencies between PVS measures reported by different methods of automated segmentation. The present review emphasizes that importance of voxel-wise evaluation of model performance, mainly with the Sørensen Dice similarity coefficient. Conventional count correlations for model validation are inadequate if the goal is to assess volumetric or morphological measures of PVS. The downside of voxel-wise evaluation is that it requires manual segmentations that require large amounts of time to produce. One possible solution is to derive these semi-automatically. Additionally, recommendations are made to facilitate rigorous development and validation of automated PVS segmentation models. In the application of automated PVS segmentation tools, publication of image quality metrics, such as the contrast-to-noise ratio, alongside descriptive statistics of PVS volumes and counts will facilitate comparability between studies. Lastly, a head-to-head comparison between two algorithms, applied to two cohorts of astronauts reveals how results can differ substantially between techniques.

Topological relationships between perivascular spaces and progression of white matter hyperintensities: A pilot study in a sample of the Lothian Birth Cohort 1936

Enlarged perivascular spaces (PVS) and white matter hyperintensities (WMH) are features of cerebral small vessel disease which can be seen in brain magnetic resonance imaging (MRI). Given the associations and proposed mechanistic link between PVS and WMH, they are hypothesized to also have topological proximity. However, this and the influence of their spatial proximity on WMH progression are unknown. We analyzed longitudinal MRI data from 29 out of 32 participants (mean age at baseline = 71.9 years) in a longitudinal study of cognitive aging, from three waves of data collection at 3-year intervals, alongside semi-automatic segmentation masks for PVS and WMH, to assess relationships. The majority of deep WMH clusters were found adjacent to or enclosing PVS (waves-1: 77%; 2: 76%; 3: 69%), especially in frontal, parietal, and temporal regions. Of the WMH clusters in the deep white matter that increased between waves, most increased around PVS (waves-1-2: 73%; 2-3: 72%). Formal statistical comparisons of severity of each of these two SVD markers yielded no associations between deep WMH progression and PVS proximity. These findings may suggest some deep WMH clusters may form and grow around PVS, possibly reflecting the consequences of impaired interstitial fluid drainage via PVS. The utility of these relationships as predictors of WMH progression remains unclear.

Computer-Aided Extraction of Select MRI Markers of Cerebral Small Vessel Disease: A Systematic Review

Cerebral small vessel disease (CSVD) is a major vascular contributor to cognitive impairment in ageing, including dementias. Imaging remains the most promising method for in vivo studies of CSVD. To replace the subjective and laborious visual rating approaches, emerging studies have applied state-of-the-art artificial intelligence to extract imaging biomarkers of CSVD from MRI scans. We aimed to summarise published computer-aided methods for the examination of three imaging biomarkers of CSVD, namely cerebral microbleeds (CMB), dilated perivascular spaces (PVS), and lacunes of presumed vascular origin. Seventy classical image processing, classical machine learning, and deep learning studies were identified. Transfer learning and weak supervision techniques have been applied to accommodate the limitations in the training data. While good performance metrics were achieved in local datasets, there have not been generalisable pipelines validated in different research and/or clinical cohorts. Future studies could consider pooling data from multiple sources to increase data size and diversity, and evaluating performance using both image processing metrics and associations with clinical measures.

Direct Rating Estimation of Enlarged Perivascular Spaces (EPVS) in Brain MRI Using Deep Neural Network

In this article, we propose a deep-learning-based estimation model for rating enlarged perivascular spaces (EPVS) in the brain’s basal ganglia region using T2-weighted magnetic resonance imaging (MRI) images. The proposed method estimates the EPVS rating directly from the T2-weighted MRI without using either the detection or the segmentation of EVPS. The model uses the cropped basal ganglia region on the T2-weighted MRI. We formulated the rating of EPVS as a multi-class classification problem. Model performance was evaluated using 96 subjects’ T2-weighted MRI data that were collected from two hospitals. The results show that the proposed method can automatically rate EPVS—demonstrating great potential to be used as a risk indicator of dementia to aid early diagnosis.

MRI-visible dilated perivascular spaces in healthy young adults: A twin heritability study

We investigated the narrow‐sense heritability of MRI‐visible dilated perivascular spaces (dPVS) in healthy young adult twins and nontwin siblings (138 monozygotic, 79 dizygotic twin pairs, and 133 nontwin sibling pairs; 28.7 ± 3.6 years) from the Human Connectome Project. dPVS volumes within basal ganglia (BGdPVS) and white matter (WMdPVS) were automatically calculated on three‐dimensional T2‐weighted MRI. In univariate analysis, heritability estimates of BGdPVS and WMdPVS after age and sex adjustment were 65.8% and 90.2%. In bivariate analysis, both BGdPVS and WMdPVS showed low to moderate genetic correlations (.30–.43) but high shared heritabilities (71.8–99.9%) with corresponding regional volumes, intracranial volumes, and other regional dPVS volumes. Older age was significantly associated with larger dPVS volume in both regions even after adjusting for clinical and volumetric variables, while blood pressure was not associated with dPVS volume although there was weak genetic correlation. dPVS volume, particularly WMdPVS, was highly heritable in healthy young adults, adding evidence of a substantial genetic contribution in dPVS development and differential effect by location. Age affects dPVS volume even in young adults, while blood pressure might have limited role in dPVS development in its normal range.

Relationship Between Venules and Perivascular Spaces in Sporadic Small Vessel Diseases

Supplemental Digital Content is available in the text.

Perivascular spaces (PVS) around venules may help drain interstitial fluid from the brain. We examined relationships between suspected venules and PVS visible on brain magnetic resonance imaging.

Perivascular spaces in the brain: anatomy, physiology and pathology

Perivascular spaces include a variety of passageways around arterioles, capillaries and venules in the brain, along which a range of substances can move. Although perivascular spaces were first identified over 150 years ago, they have come to prominence recently owing to advances in knowledge of their roles in clearance of interstitial fluid and waste from the brain, particularly during sleep, and in the pathogenesis of small vessel disease, Alzheimer disease and other neurodegenerative and inflammatory disorders. Experimental advances have facilitated in vivo studies of perivascular space function in intact rodent models during wakefulness and sleep, and MRI in humans has enabled perivascular space morphology to be related to cognitive function, vascular risk factors, vascular and neurodegenerative brain lesions, sleep patterns and cerebral haemodynamics. Many questions about perivascular spaces remain, but what is now clear is that normal perivascular space function is important for maintaining brain health. Here, we review perivascular space anatomy, physiology and pathology, particularly as seen with MRI in humans, and consider translation from models to humans to highlight knowns, unknowns, controversies and clinical relevance.

Quantitative measurements of enlarged perivascular spaces in the brain are associated with retinal microvascular parameters in older community-dwelling subjects

BACKGROUND

Perivascular Spaces (PVS) become increasingly visible with advancing age on brain MRI, yet their relationship to morphological changes in the underlying microvessels remains poorly understood. Retinal and cerebral microvessels share morphological and physiological properties. We compared computationally-derived PVS morphologies with retinal vessel morphologies in older people.

METHODS

We analysed data from community-dwelling individuals who underwent multimodal brain MRI and retinal fundus camera imaging at mean age 72.55 years (SD=0.71). We assessed centrum semiovale PVS computationally to determine PVS total volume and count, and mean per-subject individual PVS length, width and size. We analysed retinal images using the VAMPIRE software suite, obtaining the Central Retinal Artery and Vein Equivalents (CRVE and CRAE), Arteriole-to-Venule ratio (AVR), and fractal dimension (FD) of both eyes. We investigated associations using general linear models, adjusted for age, gender, and major vascular risk factors.

RESULTS

In 381 subjects with all measures, increasing total PVS volume and count were associated with decreased CRAE in the left eye (volume β=-0.170, count β=-0.184, p<0.001). No associations of PVS with CRVE were found. The PVS total volume, individual width and size increased with decreasing FD of the arterioles (a) and venules (v) of the left eye (total volume: FDa β=-0.137, FDv β=-0.139, p<0.01; width: FDa β=-0.144, FDv β=-0.158, p<0.01; size: FDa β=-0.157, FDv β=-0.162, p<0.01).

CONCLUSIONS

Increase in PVS number and size visible on MRI reflect arteriolar narrowing and lower retinal arteriole and venule branching complexity, both markers of impaired microvascular health. Computationally-derived PVS metrics may be an early indicator of failing vascular health and should be tested in longitudinal studies.

Computational quantification of brain perivascular space morphologies: Associations with vascular risk factors and white matter hyperintensities. A study in the Lothian Birth Cohort 1936

BACKGROUND AND PURPOSE

Perivascular Spaces (PVS), also known as Virchow-Robin spaces, seen on structural brain MRI, are important fluid drainage conduits and are associated with small vessel disease (SVD). Computational quantification of visible PVS may enable efficient analyses in large datasets and increase sensitivity to detect associations with brain disorders. We assessed the associations of computationally-derived PVS parameters with vascular factors and white matter hyperintensities (WMH), a marker of SVD.

PARTICIPANTS

Community dwelling individuals (n = 700) from the Lothian Birth Cohort 1936 who had multimodal brain MRI at age 72.6 years (SD = 0.7).

METHODS

We assessed PVS computationally in the centrum semiovale and deep corona radiata on T2-weighted images. The computationally calculated measures were the total PVS volume and count per subject, and the mean individual PVS length, width and size, per subject. We assessed WMH by volume and visual Fazekas scores. We compared PVS visual rating to PVS computational metrics, and tested associations between each PVS measure and vascular risk factors (hypertension, diabetes, cholesterol), vascular history (cardiovascular disease and stroke), and WMH burden, using generalized linear models, which we compared using coefficients, confidence intervals and model fit.

RESULTS

In 533 subjects, the computational PVS measures correlated positively with visual PVS ratings (PVS count r = 0.59; PVS volume r = 0.61; PVS mean length r = 0.55; PVS mean width r = 0.52; PVS mean size r = 0.47). PVS size and width were associated with hypertension (OR 1.22, 95% CI [1.03 to 1.46] and 1.20, 95% CI [1.01 to 1.43], respectively), and stroke (OR 1.34, 95% CI [1.08 to 1.65] and 1.36, 95% CI [1.08 to 1.71], respectively). We found no association between other PVS measures and diabetes, hypercholesterolemia or cardiovascular disease history. Computational PVS volume, length, width and size were more strongly associated with WMH (PVS mean size versus WMH Fazekas score β = 0.66, 95% CI [0.59 to 0.74] and versus WMH volume β = 0.43, 95% CI [0.38 to 0.48]) than computational PVS count (WMH Fazekas score β = 0.21, 95% CI [0.11 to 0.3]; WMH volume β = 0.14, 95% CI [0.09 to 0.19]) or visual score. Individual PVS size showed the strongest association with WMH.

CONCLUSIONS

Computational measures reflecting individual PVS size, length and width were more strongly associated with WMH, stroke and hypertension than computational count or visual PVS score. Multidimensional computational PVS metrics may increase sensitivity to detect associations of PVS with risk exposures, brain lesions and neurological disease, provide greater anatomic detail and accelerate understanding of disorders of brain fluid and waste clearance.

Perivascular spaces and their associations with risk factors, clinical disorders and neuroimaging features: A systematic review and meta-analysis

BACKGROUND

Perivascular spaces, visible on brain magnetic resonance imaging, are thought to be associated with small vessel disease, neuroinflammation, and to be important for cerebral hemodynamics and interstitial fluid drainage.

AIMS

To benchmark current knowledge on perivascular spaces associations with risk factors, neurological disorders, and neuroimaging lesions, using systematic review and meta-analysis.

SUMMARY OF REVIEW

We searched three databases for perivascular spaces publications, calculated odds ratios with 95% confidence interval and performed meta-analyses to assess adjusted associations with perivascular spaces. We identified 116 relevant studies (n = 36,108) but only 23 (n = 12,725) were meta-analyzable. Perivascular spaces assessment, imaging and clinical definitions varied. Perivascular spaces were associated (n; OR, 95%CI, p) with ageing (8395; 1.47, 1.28-1.69, p = 0.00001), hypertension (7872; 1.67, 1.20-2.31, p = 0.002), lacunes (4894; 3.56, 1.39-9.14, p = 0.008), microbleeds (5015; 2.26, 1.04-4.90, p = 0.04) but not WMH (4974; 1.54, 0.71-3.32, p = 0.27), stroke or cognitive impairment. There was between-study heterogeneity. Lack of appropriate data on other brain disorders and demographic features such as ethnicity precluded analysis.

CONCLUSIONS

Despite many studies, more are required to determine potential pathophysiological perivascular spaces involvement in cerebrovascular, neurodegenerative and neuroinflammatory disorders.

Longitudinal serum S100β and brain aging in the Lothian Birth Cohort 1936

Elevated serum and cerebrospinal fluid concentrations of S100β, a protein predominantly found in glia, are associated with intracranial injury and neurodegeneration, although concentrations are also influenced by several other factors. The longitudinal association between serum S100β concentrations and brain health in nonpathological aging is unknown. In a large group (baseline N = 593; longitudinal N = 414) of community-dwelling older adults at ages 73 and 76 years, we examined cross-sectional and parallel longitudinal changes between serum S100β and brain MRI parameters: white matter hyperintensities, perivascular space visibility, white matter fractional anisotropy and mean diffusivity (MD), global atrophy, and gray matter volume. Using bivariate change score structural equation models, correcting for age, sex, diabetes, and hypertension, higher S100β was cross-sectionally associated with poorer general fractional anisotropy (r = -0.150, p = 0.001), which was strongest in the anterior thalamic (r = -0.155, p < 0.001) and cingulum bundles (r = -0.111, p = 0.005), and survived false discovery rate correction. Longitudinally, there were no significant associations between changes in brain imaging parameters and S100β after false discovery rate correction. These data provide some weak evidence that S100β may be an informative biomarker of brain white matter aging.

Perivascular Spaces Segmentation in Brain MRI Using Optimal 3D Filtering

Perivascular Spaces (PVS) are a feature of Small Vessel Disease (SVD), and are an important part of the brain’s circulation and glymphatic drainage system. Quantitative analysis of PVS on Magnetic Resonance Images (MRI) is important for understanding their relationship with neurological diseases. In this work, we propose a segmentation technique based on the 3D Frangi filtering for extraction of PVS from MRI. We used ordered logit models and visual rating scales as alternative ground truth for Frangi filter parameter optimization and evaluation. We optimized and validated our proposed models on two independent cohorts, a dementia sample (N = 20) and patients who previously had mild to moderate stroke (N = 48). Results demonstrate the robustness and generalisability of our segmentation method. Segmentation-based PVS burden estimates correlated well with neuroradiological assessments (Spearman’s ρ = 0.74, p < 0.001), supporting the potential of our proposed method.

MR Imaging-based Multimodal Autoidentification of Perivascular Spaces (mMAPS): Automated Morphologic Segmentation of Enlarged Perivascular Spaces at Clinical Field Strength

Purpose To describe a fully automated segmentation method that yields object-based morphologic estimates of enlarged perivascular spaces (ePVSs) in clinical-field-strength (3.0-T) magnetic resonance (MR) imaging data. Materials and Methods In this HIPAA-compliant study, MR imaging data were obtained with a 3.0-T MR imager in research participants without dementia (mean age, 85.3 years; range, 70.4-101.2 years) who had given written informed consent. This method is built on (a) relative normalized white matter, ventricular and cortical signal intensities within T1-weighted, fluid-attenuated inversion recovery, T2-weighted, and proton density data and (b) morphologic (width, volume, linearity) characterization of each resultant cluster. Visual rating was performed by three raters, including one neuroradiologist, after established single-section guidelines. Correlations between visual counts and automated counts, as well session-to-session correlation of counts within each participant, were assessed with the Pearson correlation coefficient r. Results There was a significant correlation between counts by visual raters and automated detection of ePVSs in the same section (r = 0.65, P < .001; r = 0.69, P < .001; and r = 0.54, P < .01 for raters 1, 2, and 3, respectively). With regard to visual ratings and whole-brain count consistency, average visual rating scores were highly correlated with automated detection of total burden volume (r = 0.58, P < .01) and total ePVS number (r = 0.76, P < .01). Morphology of clusters across 28 data sets was consistent with published radiographic estimates of ePVS; mean width of clusters segmented was 3.12 mm (range, 1.7-13.5 mm). Conclusion This MR imaging-based method for multimodal autoidentification of perivascular spaces yields individual whole-brain morphologic characterization of ePVS in clinical MR imaging data and is an important tool in the detailed assessment of these features. ^© RSNA, 2017 Online supplemental material is available for this article.

Reliability of an automatic classifier for brain enlarged perivascular spaces burden and comparison with human performance

In the brain, enlarged perivascular spaces (PVS) relate to cerebral small vessel disease (SVD), poor cognition, inflammation and hypertension. We propose a fully automatic scheme that uses a support vector machine (SVM) to classify the burden of PVS in the basal ganglia (BG) region as low or high. We assess the performance of three different types of descriptors extracted from the BG region in T2-weighted MRI images: (i) statistics obtained from Wavelet transform's coefficients, (ii) local binary patterns and (iii) bag of visual words (BoW) based descriptors characterizing local keypoints obtained from a dense grid with the scale-invariant feature transform (SIFT) characteristics. When the latter were used, the SVM classifier achieved the best accuracy (81.16%). The output from the classifier using the BoW descriptors was compared with visual ratings done by an experienced neuroradiologist (Observer 1) and by a trained image analyst (Observer 2). The agreement and cross-correlation between the classifier and Observer 2 (κ = 0.67 (0.58-0.76)) were slightly higher than between the classifier and Observer 1 (κ = 0.62 (0.53-0.72)) and comparable between both the observers (κ = 0.68 (0.61-0.75)). Finally, three logistic regression models using clinical variables as independent variable and each of the PVS ratings as dependent variable were built to assess how clinically meaningful were the predictions of the classifier. The goodness-of-fit of the model for the classifier was good (area under the curve (AUC) values: 0.93 (model 1), 0.90 (model 2) and 0.92 (model 3)) and slightly better (i.e. AUC values: 0.02 units higher) than that of the model for Observer 2. These results suggest that, although it can be improved, an automatic classifier to assess PVS burden from brain MRI can provide clinically meaningful results close to those from a trained observer.