Elsevier

NeuroImage

Volume 165, 15 January 2018, Pages 294-305
NeuroImage

Combining images and anatomical knowledge to improve automated vein segmentation in MRI

https://doi.org/10.1016/j.neuroimage.2017.10.049Get rights and content

Abstract

Purpose

To improve the accuracy of automated vein segmentation by combining susceptibility-weighted images (SWI), quantitative susceptibility maps (QSM), and a vein atlas to produce a resultant image called a composite vein image (CV image).

Method

An atlas was constructed in common space from manually traced MRI images from ten volunteers. The composite vein image was derived for each subject as a weighted sum of three inputs; an SWI image, a QSM image and the vein atlas. The weights for each input and each anatomical location, called template priors, were derived by assessing the accuracy of each input over an independent data set. The accuracy of vein segmentations derived automatically from each of the CV image, SWI, and QSM image sets was assessed by comparison with manual tracings. Three different automated vein segmentation techniques were used, and ten performance metrics evaluated.

Results

Vein segmentations using the CV image were comprehensively better than those derived from SWI or QSM images (mean Cohen's d = 1.1). Sixty permutations of performance metric, benchmark image, and automated segmentation technique were evaluated. Vein identification improvements that were both large and significant (Cohen's d > 0.80, p < 0.05) were found in 77% of the permutations, compared to no improvement in 5%.

Conclusion

The accuracy of automated vein segmentations derived from the composite vein image was overwhelmingly superior to segmentations derived from SWI or QSM alone.

Introduction

Mapping cerebral veins using magnetic resonance (MR) images has until recently been technically challenging. Cerebral venograms are increasingly important for advancing our knowledge of cerebral vascularisation, oxygenation, metabolism, and studies of cerebrovascular topology. The use of venograms in clinical research applications is growing rapidly, including for quantifying oxygen saturation (Fan et al., 2014), measuring the metabolic rate of oxygen consumption (Rodgers et al., 2016), analyzing possible fMRI confounders (Vigneau-Roy et al., 2014), and planning neurosurgery (Grabner et al., 2017).

Traditional vein imaging techniques require invasive contrast agents, have potential arterial confounds, and are limited to the large vessels, due to the reduced volume and flow of smaller cerebrovasculature segments. However, magnetic susceptibility provides an intrinsic contrast mechanism that is exquisitely sensitive to the presence of iron, particularly deoxygenated iron-rich haemoglobin proteins within red blood cells, making susceptibility techniques very useful for imaging small as well as large veins. The magnetic susceptibility of blood is modulated by oxygen (Pauling and Coryell, 1936), which facilitates the separation of arteries and veins, whilst providing a mechanism to quantify oxygen saturation (Fan et al., 2014).

Susceptibility-weighted imaging (SWI) and quantitative susceptibility mapping (QSM) are MR techniques based on magnetic susceptibility that provide a non-invasive method of imaging the cerebral veins. QSM and SWI derive contrast from gradient-recalled echo (GRE) phase information and have been applied to stroke, multiple sclerosis, cerebrovascular disease, and examined in clinical and preclinical studies (Betts et al., 2016, Deistung et al., 2016, Fan et al., 2015, Fujima et al., 2011, Goodwin et al., 2015, Jain et al., 2010, Li et al., 2013, Liu and Li, 2016, Rodgers et al., 2013, Santhosh et al., 2009). The way in which SWI and QSM process the phase information is very different.

SWI multiplies a non-linear mapping of high-pass filtered GRE phase with the GRE magnitude image, compounding the effects of signal cancellation from incoherent signals within each voxel and phase accumulation due to local sources of magnetic susceptibility (Haacke et al., 2004). Non-local sources are also included, such as the extravascular phase information, resulting in the magnification of small veins. The presentation of non-local sources, and non-linear mapping, generates a non-quantitative image best suited to radiological interpretation.

QSM estimates the magnetic susceptibility distribution directly by inverting the magnetic field information captured in the phase image (Marques and Bowtell, 2005, Salomir et al., 2003). Mathematically, QSM involves a linear system inversion that is ill-posed and requires regularization or fitting (Li et al., 2015, Liu et al., 2017, Wang and Liu, 2015, Wharton et al., 2010). QSM has the benefits of being quantitative and is designed to resolve extravascular field effects, leaving only local sources of magnetic susceptibility contrast.

The differing approaches (QSM and SWI) have unique image contrasts, and each have their own vein-like confounders. SWI images, for instance, do not distinguish between signal cancellation due to venous blood, and low concentrations of free protons (Haacke et al., 2004). The lack of distinction is problematic when analyzing veins which reside near non-vein low signal structures, such as in the vicinity of the tentorium and in the interhemispheric fissure (due to the falx cerebri). Both SWI and QSM also suffer different artefacts, such as cruciform artefacts in QSM images. Additionally, vein contrast in both SWI and QSM is reduced by iron depositions. The high amount of iron, for instance in the basal ganglia, can impair venous segmentation.

As neither QSM nor SWI isolate blood signal intrinsically, unlike spin-labelling or contrast agent-based techniques, venous voxels within the brain must be identified before the veins can be analysed. The process of identifying venous voxels in the brain, or segmentation, produces a vein mask that can then be used to extract the vein signal from an image, or examined directly for topographic analysis. In this work, the term segmentation is used as a noun to refer to a binary mask that labels each voxel as vein or non-vein.

A number of algorithms for automatic segmentation of blood vessels in the body have been proposed, including shape-driven, intensity-driven and hybrid approaches (Lesage et al., 2009). A common approach in the analysis of SWI and QSM data is to employ a preliminary filtering step, such as Hessian-based filtering (Frangi et al., 1998), before applying a simple threshold classification method (Vigneau-Roy et al., 2014). Hessian-based filtering and thresholding has been used to construct a vein-atlas to study multiple-sclerosis (Grabner et al., 2014). These filtering techniques have also been applied to build vascular network models using both QSM (Kociński et al., 2017) and time-of-flight images (Hsu et al., 2017). Recent work has combined Hessian-based filtering into a segmentation framework with diffusion techniques to overcome noise and low vein visibility (Bazin et al., 2016, Manniesing et al., 2006). Statistical modeling of spatial relations has also been proposed to improve continuity and smoothness in vein segmentation (Bériault et al., 2014, Ward et al., 2017b).

The previously mentioned work focused upon SWI (Bériault et al., 2014, Vigneau-Roy et al., 2014) or QSM (Bazin et al., 2016, Kociński et al., 2017, Ward et al., 2017b), and did not attempt to extract information from both images. Methods have been proposed that merge SWI with QSM (Ward et al., 2015), SWI with R2* maps (Monti et al., 2015), and QSM with both SWI and R2* (Deistung et al., 2013). These approaches were globally homogeneous, i.e., they combined voxel intensities without consideration for anatomical location. As SWI and QSM have differing image contrasts, and artefacts that are specific to anatomy, it is possible that an improved segmentation could be achieved if the method for combining the two images was sensitive to spatial location.

Prior anatomical knowledge has recently been incorporated into a vein segmentation technique to reduce false positives in specific brain regions (Bériault et al., 2015). However, this approach was limited to specific deep-brain regions (particularly the basal ganglia), it did not directly address boundaries between tissue types and neural structures, and it was hand-tuned.

There are two anatomical factors that contribute to vein segmentation accuracy. The first is vein anatomy, i.e., expected vein occurrence, size and shape at an anatomical location. The second is image contrast, i.e., expected tissue signal relative to vein signal, which is specific to SWI and QSM. In this study, these two factors are exploited to improve cerebrovenous contrast and subsequent vein segmentation accuracy. We propose a vein identification and segmentation method that is based on a locally varying combination of SWI and QSM contrast which is informed by known vein anatomy in specific neuroanatomical structures. The proposed method derives a single composite vein image (CV image) that incorporates the strengths of SWI and QSM, with the anatomical knowledge of a vein atlas.

The CV image is generated from three input images (SWI, QSM and atlas) that are combined using a weighted-sum. The weights are derived from template priors that capture the location-specific venous contrast of the three input images throughout the brain. Separate vein atlases and template priors were calculated for each subject within the study from an independent sample of the cohort to ensure data independence. Future applications of the technique would use a single template prior and atlas calculated from the entire cohort. The CV image was compared to SWI and QSM images for the purpose of vein segmentation using automated techniques. Segmented CV images were compared with segmented SWI and QSM images using a broad array of accuracy measures and three automated segmentation techniques.

Section snippets

Methods

All procedures were reviewed and approved by the local ethics committee. Informed consent was obtained from all volunteers. The code and data used in this study have been made available to the public using GitHub and figshare respectively (https://doi.org/10.4225/03/57B6AB25DDBDC) (Ward, 2017a, Ward et al., 2017c).

Results

A visual inspection of a single CV image (see Fig. 3) shows the strength of combining the three inputs. The transverse slice (Fig. 3F–J) depicts the iron rich basal ganglia, which are particularly bright on QSM (Fig. 3G). The CV image is able to partially suppress the basal ganglia, whilst retaining the high vein contrast of QSM. The atlas-free CV image (Fig. 3D, I and N) does not show the same level of suppression in the basal ganglia as the CV image. The vein atlas (Fig. 3C, H and M) has been

Discussion

In this work, a composite vein (CV) imaging technique was proposed that combined three sources of vein information, an atlas, an SWI image, and a QSM image. The CV image showed a large improvement in vein segmentation accuracy when compared with SWI and QSM images. A robust improvement was observed in the majority of permutations across ten performance metrics and three segmentation methods.

The CV image was found to combine the complementary strengths of SWI and QSM, and produce an image with

Conclusions

A large improvement in vein segmentation accuracy was achieved using the composite vein image technique. The composite vein image technique incorporates the heterogeneous vein contrast profile across the brain to extract the complementary information available from SWI and QSM images, and a vein atlas. The technique's performance was evaluated with multiple segmentation techniques and metrics. The accuracy provided by the composite vein image allows improved quantification of cerebrovenous

Acknowledgements

We thank the investigators of the ASPREE study and the ASPREE-NEURO sub-study for the coordination and recruitment of the volunteers. We thank the volunteers for their time and willingness to participate.

The Alzheimer's Australia Dementia Research Foundation (AADRF) the Victorian Life Sciences Computation Initiative (VLSCI), the Multi-model Australian Sciences Imaging and Visualisation Environment (MASSIVE) and the National Health and Medical Research Council (NHMRC) supported this work (NHMRC

References (66)

  • Y. Xu et al.

    The role of voxel aspect ratio in determining apparent vascular phase behavior in susceptibility weighted imaging

    Magn. Reson. Imaging

    (2006)
  • P.L. Bazin et al.

    Vessel segmentation from quantitative susceptibility maps for local oxygenation venography

  • S. Bériault et al.

    Automatic markov random field segmentation of susceptibility-weighted MR venography

  • S. Bériault et al.

    Automatic SWI venography segmentation using conditional random fields

    IEEE Trans. Med. Imaging

    (2015)
  • J. Cohen

    Statistical Power Analysis for the Behavioral Sciences

    (2013)
  • A. Deistung et al.

    Susceptibility weighted imaging at ultra high magnetic field strengths: theoretical considerations and experimental results

    Magn. Reson. Med.

    (2008)
  • A. Deistung et al.

    Overview of quantitative susceptibility mapping

    NMR Biomed.

    (2016)
  • A. Deistung et al.

    Optimal enhancement of brain structures by combining different MR contrasts: demonstration of venous vessel enhancement in multi-echo gradient-echo MRI

  • A.P. Dempster et al.

    Maximum likelihood from incomplete data via the EM algorithm

    J. R. Stat. Soc. Ser. B Methodol.

    (1977)
  • L.R. Dice

    Measures of the amount of ecologic association between species

    Ecology

    (1945)
  • D.L. Dowe

    Foreword re C. S. Wallace

    Comput. J.

    (2008)
  • T. Drew et al.

    The invisible gorilla strikes again: sustained inattentional blindness in expert observers

    Psychol. Sci.

    (2013)
  • A.P. Fan et al.

    Quantitative oxygenation venography from MRI phase

    Magn. Reson. Med.

    (2014)
  • A.P. Fan et al.

    Quantitative oxygen extraction fraction from 7-Tesla MRI phase: reproducibility and application in multiple sclerosis

    J. Cereb. Blood Flow. Metab.

    (2015)
  • A.F. Frangi et al.

    Multiscale vessel enhancement filtering

  • G. Gerig et al.

    Valmet: a new validation tool for assessing and improving 3D object segmentation

  • I.J. Good

    Rational decisions

    J. R. Stat. Soc. Ser. B Methodol.

    (1952)
  • J.A. Goodwin et al.

    Susceptibility-weighted phase imaging and oxygen extraction fraction measurement during sedation and sedation recovery using 7T MRI

    J. Neuroimaging

    (2015)
  • W.J. Goscinski et al.

    The multi-modal Australian ScienceS Imaging and Visualization Environment (MASSIVE) high performance computing infrastructure: applications in neuroscience and neuroinformatics research. Front

    Neuroinformatics

    (2014)
  • G. Grabner et al.

    Group specific vein-atlasing: an application for analyzing the venous system under normal and multiple sclerosis conditions

    J. Magn. Reson. Imaging

    (2014)
  • G. Grabner et al.

    Post mortem validation of MRI-identified veins on the surface of the cerebral cortex as potential landmarks for neurosurgery

    Front. Neurosci.

    (2017)
  • A.I. Group

    Study design of ASPirin in Reducing Events in the Elderly (ASPREE): a randomized, controlled trial

    Contemp. Clin. Trials

    (2013)
  • E.M. Haacke et al.

    Susceptibility weighted imaging (SWI)

    Magn. Reson Med.

    (2004)
  • Cited by (25)

    • Individual differences in haemoglobin concentration influence bold fMRI functional connectivity and its correlation with cognition

      2020, NeuroImage
      Citation Excerpt :

      This was not the case in the female group, where the distribution of the linear coefficients between haemoglobin and functional connectivity edge-weights overlapped the null distribution (Fig. 2A). To determine if the haemoglobin influence on functional connectivity values was related to the proximity to cerebral veins, a map of the highest haemoglobin functional connectivity associations was compared with an atlas of the cerebral veins (Ward et al., 2018). Fig. 3 shows the spatial map of the top 10% t coefficients, and the probabilistic location of the major draining veins (Ward et al., 2018).

    • A critical assessment of data quality and venous effects in sub-millimeter fMRI

      2019, NeuroImage
      Citation Excerpt :

      For a more intuitive visualization of the vasculature, we performed an analysis in which we calculate the minimum intensity observed in the EPI data over a 1-cm slab positioned in occipital cortex (Fig. 7B, lower left). This minimum intensity projection analysis (Haacke et al., 2009; Ward et al., 2018) reveals the branching, tree-like structure of the vasculature (Fig. 7B, top row, 2nd column) and is consistent with the results of the same analysis applied to a susceptibility weighted imaging (SWI) scan acquired at 0.4-mm resolution (Fig. 7B, top row, 1st column). To understand how results manifest at more standard fMRI resolutions, we repeated the minimum intensity projection analysis for simulated low-resolution 2.4-mm EPI data, obtained by spatially smoothing the high-resolution 7 T EPI data (Fig. 7B, top row, 3rd column), as well as actual low-resolution 2.4-mm 3 T EPI data (Fig. 7B, top row, 4th column).

    View all citing articles on Scopus
    View full text