2019 Research Forum

Automated and Standardized Quantification of Mild Cerebral Small Vessel Disease in Computed Tomography

Ronald Crandall MS IV, Sudha Challa MD

Conclusions

Visual Scale vs. Volumetric Analysis

Cerebral Small Vessel Disease Cerebral small vessel disease is commonly seen by white matter change (WMC) on brain imaging. These findings are thought to be due to incomplete ischemic change from cerebral blood flow regulation failure due to atherosclerotic disease and breakdown of the blood-brain barrier. 1 Introduction Choice of Imaging WMCs are best seen as periventricular hyperintensities on the MRI FLAIR (fluid-attenuated inversion recovery) sequence. As computed tomography (CT) is less costly and more rapidly performed, it is frequently the initial imaging study ordered in the neurological workup of a patient. WMCs are seen as hypodensities on computed tomography. Mild WMCs are often difficult to detect on computed tomography due to the narrow densities measured in Hounsfield units (HU) of the brain parenchyma.

Even though CT images are less sensitive than MRI, CTs are utilized more frequently, performed more quickly and less costly. Automated detection of WMCs on computed tomography would be beneficial. Mild WMCs on CT are difficult to interpret and in practice are not graded with the Age-Related White Matter Change scale by Wahlund et al. 7

Fazekas scale , most commonly used, first proposed in 1987. 2 • 0-3 scale primarily of deep WMC.

Age-related white matter changes scale (ARWMC) • validated for both MRI and CT, and is correlated well with cognitive impairment • 0-3 scale of both periventricular WMC and deep WMC. Visual scales do not track volumetric changes , and so have difficulty assessing WMCs over time. Manual volumetric analysis is extremely time-consuming . While automated approaches for volumetric analysis by MRI have been developed, very few are openly available for CT. Therefore we developed an automated algorithm to detect WMC volumes.

1

2

3

In our study, similar to Hanning et al., total white matter density correlated with overall WMCs. 8

Cases courtesy of Dr Bruno Muzio,Radiopaedia.org

Methods

Our algorithm identifies separation between periventricular and deep white matter changes as well as provides a useful visual aid in the 3 rd , 4 th and 5 th decades.

For our research purpose, we performed a retrospective analysis of previously identified individuals with normal or white matter changes. Brain extraction and registration using a CT template as described in the methods by Muschelli et al. 3 CT images were subsequently tissue-type segmented according to the methods described in Cauley et al. utilizing FSL (FMRIB software library). 4,5 We utilized 3 instead of 2 tissue-type classes as this achieved better delineation between WM and partial volume averaging from CSF. The brain images were then further segmented between 1-30 HU densities. Neighborhood voxels were selected between 1-20 HU densities utilizing the neighboring method described by Muschelli et al. originally for intracranial hemorrhage. 6 Mean neighborhood densities were calculated using a 3x3mm voxel box (voxel = volume). Periventricular voxels were then used to create masks of the WM, revealing adjacent voxels in the deep white matter. The two series below show volumetric highlighting of hypodense areas in a mild and severe case.

An age-associated trend in increasing WMCs in normal individuals was noted in both periventricular and deep white matter.

This algorithm may allow CT to be utilized to identify WMCs and their severity accurately.

Discussion

Further validation of the methodology in patients without trauma, chronic infection, or history of polysubstance abuse is warranted.

White Matter Segmentation

Total WMC neighborhood

Deep WMC neighborhood

Periventricular WMC neighborhood

Brain Extraction

CT

Mild

References 1. Xiong, YunYun,andVincentMok. “Age-RelatedWhite Matter Changes.” JournalofAgingResearch,vol.2011, Aug. 2011,p. 617927, doi:10.4061/2011/617927. 2 . Fazekas, F., etal. “MR Signal Abnormalitiesat 1.5 T inAlzheimer’sDementiaandNormalAging.” AJR.American JournalofRoentgenology, vol.149,no. 2,Aug. 1987,pp.351–56, doi:10.2214/ajr.149.2.351. 3 .Muschelli, John, Natalie L.Ullman,etal. “ValidatedAutomatic Brain Extraction ofHeadCT Images.”NeuroImage, vol.114, July2015,pp. 379–85,doi:10.1016/j.neuroimage.2015.03.074. 4 .Cauley,Keith A., etal. “Automated Segmentation ofHeadComputed Tomography ImagesUsing FSL.” Journalof ComputerAssisted Tomography, vol.42,no. 1, 2018,pp.104–10, doi:10.1097/RCT.0000000000000660. 5 . Jenkinson, Mark,etal. “FSL.”NeuroImage,vol.62,no. 2,Aug. 2012,pp.782–90, doi:10.1016/j.neuroimage.2011.09.015. 6 .Muschelli, John, etal. “Neuroconductor: AnRPlatform forMedical ImagingAnalysis.”Biostatistics , vol.20,no. 2, Apr. 2019,pp. 218–39,doi:10.1093/biostatistics/kxx068. 7 .Wahlund, L.O.,etal. “ANewRating Scale forAge-RelatedWhite Matter Changes Applicable to MRIandCT.” Stroke; a JournalofCerebralCirculation, vol.32,no. 6, June 2001,pp.1318–22, https://www.ncbi.nlm.nih.gov/pubmed/11387493. 8 .Hanning, Uta, etal. “Quantitative RapidAssessmentof Leukoaraiosis inCT :Comparison to Gold Standard MRI.” ClinicalNeuroradiology, vol.29,no. 1,Mar. 2019,pp. 109–15,doi:10.1007/s00062-017-0636-2. Imagescreated utilizingMango fromResearch Imaging Institute, UTHSCSA, FSL EYES,and GIMP. Acknowledgments My mentor Dr. Sudha Challa for her continued support, devoted time and patience. Multiple file room staff who helped make this study possible and of course our patients. The numerous authors and editors of the software packages utilized in this study. Studies with extensive artifact and severe traumatic brain injury were excluded from the analysis. For further research, the application of the above techniques to a larger population for additional training and study with co- morbid conditions may increase more accurate detection. Additional algorithm modification such as sub-cortical masking may better reveal pathologically distinct disease processes. A further modification to better segment WM loss near CSF would also be useful, but challenging. In the future with the addition of more data, machine learning algorithms would be of further benefit. Clinical use of automated detection of WMCs will help standardize initial to severe disease stages by providing both a visual aid and volumetric data. The addition of volumetric analysis provides for better longitudinal description of WMCs.

CT

FLAIR

Periventricular vs Deep WMCs Initial periventricular WMCs are pathologically distinct from irregular periventricular WMC and confluent deep WMC changes. Initial periventricular changes have primarily partial myelin loss without atherosclerotic changes Irregular or confluent WMCs show myelin rarefaction with or without incomplete parenchymal loss resembling infarction.

Severe

Periventricular & Isolated Deep WMC

FLAIR

Results

CT

Irregular Periventricular & Confluent Deep WMC

Age Group

Total WM Volume

Total WM Density

Total WMC Volume Periventricular WMC Volume Deep WMC Volume

20-29 normal (7)

611.48 651.64

33.10 31.29 0.037 33.66 30.85 <0.001 33.02 31.53

1.95

0.71 3.72 0.001 1.25 6.80 0.005 2.18 5.97 0.017

0.50 2.83

20-29 with WMCs (6)

18.17 0.004

P values

0.155

<0.001

30-39 normal (12)

655.26 643.44

4.15

0.62 4.93 0.008 1.20 4.15 0.033

30-39 with WMCs (10)

40.05 0.016

P values

0.310

40-49 normal (9)

676.06 643.45

4.87

40-49 with WMCs (11)

36.40 0.016

FLAIR

CT

P values

0.062

0.024

* volume is in cm 3 , density is inHounsfield Units

127