Evaluation of three automatic brain vessel segmentation methods for stereotactical trajectory planning

https://doi.org/10.1016/j.cmpb.2019.105037Get rights and content

Abstract

Background and objective

Stereotactical procedures require exact trajectory planning to avoid blood vessels in the trajectory path. Innovation in imaging and image recognition techniques have facilitated the automatic detection of blood vessels during the planning process and may improve patient safety in the future.

To assess the feasibility of a vessel detection and warning system using currently available imaging and vessel segmentation techniques.

Methods

Image data were acquired from post-contrast, isovolumetric T1-weighted sequences (T1CE) and time.-of-flight MR angiography at 3T or 7T from a total of nine subjects. Vessel segmentation by a combination of a vessel-enhancement filter with subsequent level-set segmentation was evaluated using three different methods (Vesselness, FastMarching and LevelSet) in 45 stereotactic trajectories. Segmentation results were compared to a gold-standard of manual segmentation performed jointly by two human experts.

Results

The LevelSet method performed best with a mean interclass correlation coefficient (ICC) of 0.76 [0.73, 0.81] compared to the FastMarching method with ICC 0.70 [0.67, 0.73] respectively. The Vesselness algorithm achieved clearly inferior overall performance with a mean ICC of 0.56 [0.53, 0.59]. The differences in mean ICC between all segmentation methods were statistically significant (p < 0.001 with post-hoc p < 0.026). The LevelSet method performed likewise good in MPRAGE and 3T-TOF images and excellent in 7T-TOF image data. The negative predictive value (NPV) was very high (>97%) for all methods and modalities. Positive predictive values (PPV) were found in the overall range of 65–90% likewise depending on algorithm and modality. This pattern reflects the disposition of all segmentation methods – in case of misclassification - to produce preferentially false-positive than false-negative results. In a clinical setting, two to three potential collision warnings would be given per trajectory on average with a PPV of around 50%.

Conclusions

It is feasible to integrate a clinically meaningful vessel detection and collision warning system into stereotactical planning software. Both, T1CE and MRA sequences are suitable as image data for such an application.

Introduction

Stereotactical procedures such as brain biopsy (frame-based and frameless) or placement of electrodes for deep-brain stimulation (DBS) require exact trajectory planning to maximize effectiveness while minimizing morbidity and mortality. Consequently, avoidance of blood vessels in the trajectory path constitutes an indispensable part of procedure planning to prevent bleeding and ischemic complications.

Steady innovation in magnetic resonance imaging (MRI) is leading to continuously increasing spatial resolution and image contrast, whereby more and more smaller vessels can be visualized, making trajectory planning, which is still performed manually by examination of multi-planar reconstructions of image data, even more complex. Consequently, there is a growing demand for advanced stereotactical planning tools. These software systems should be capable of detecting blood vessels in the trajectory path and guide the surgeon to stay clear of such vessels.

Robust vessel segmentation is a cornerstone of any such planning system and many vessel enhancement algorithms have been evaluated using time-of-flight (TOF) magnetic resonance angiography (MRA) or CT angiography (CTA) images [1], [2], [3].

Although both imaging modalities are vessel-specific and thus arguably represent the best source imaging for vessel detection, they require time (MRA) or considerable radiation and contrast-agent exposure (CTA). These apparent drawbacks may explain the fact that such imaging protocols are not part of routine stereotactical imaging and that vessel detection is not part of contemporary stereotactical planning software yet.

Conversely, contrast-enhanced T1-weighted (T1CE) MRI sequences are an indispensable component of any MRI used in stereotactical planning. Given that this data is readily available for almost every stereotactic procedure, the question arises if robust vessel detection is possible with this imaging modality.

We have developed a software application that incorporates three vessel segmentation algorithms to detect brain vessels from T1CE and MRA images. The system is designed to generate warnings if minimum-distance constraints between vasculature and trajectory are violated. The software application was developed in C++ using the Qt Toolkit, The Visualization Toolkit (VTK) [4] and the Insight Segmentation and Registration Toolkit (ITK) [5], [6].

Vessel segmentation was performed by the combination of a generalized multiscale objectness filter [7] (derived from the vesselness filter of Frangi [8]) with subsequent level-set development of the vessel boundary towards the vessel edges.

The objective of this pilot study was to evaluate the performance of this software in 45 stereotactic trajectories. Segmentation results in T1CE and 3T or 7T MRA were compared to gold-standard manual segmentation by human experts. Furthermore, the accuracy of collision warnings in a clinical setting was evaluated.

Section snippets

Study population

Six brain tumor patients (four males and two females) scheduled for stereotactic biopsy, were included in this study. Additionally, image data from three healthy volunteers (2 males and 1 female) examined using ultra-high-field 7T MRI were collected.

Ethics approval

The study was performed in accordance with the declaration of Helsinki. Design of this study, image acquisition using optimized MR sequences, collection and analysis of patient data were approved by the institutional review board of the University

Processing time

Computation time was measured on an Intel® Core i7-3770 (4 [email protected] GHz) running FreeBSD-11.2 (Table 2). As expected for the Vesselness and FastMarching methods, linear regression analysis revealed a strong linear correlation between total computation time and the number of voxels in the bounding box of the trajectory.

The application of the LevelSet method led to a three-fold increase of processing time compared to the other methods with an average of 17.6 [16.9, 18.4] seconds. Depending on

Discussion

The results of our analysis support our hypothesis that meaningful automatic vessel collision warning for stereotactic planning is feasible. The combination of a vessel enhancement filter [8], [12] with subsequent local level-set segmentation as in the FastMarching and LevelSet methods performed significantly better than vessel enhancement filtering alone.

Furthermore, we could show that dedicated vessel imaging (MRA) does not seem to be required if high-quality T1CE images are available. Many

Conclusions

From the data of our pilot study, we conclude that it is feasible to integrate a clinically meaningful vessel detection and collision warning system into stereotactical planning software. Such a system is capable of operating on either MRA or T1CE images.

Declaration of Competing Interest

All authors declare that no conflicts of interest do exist.

Acknowledgements

We thank Prof. Dr. Armin Nagel for kindly providing the 7T-TOF image data acquired at the German Cancer Research Center, Heidelberg.

References (30)

  • R. Phellan et al.

    Comparison of vessel enhancement algorithms applied to time-of-flight MRA images for cerebrovascular segmentation

    Med. Phys.

    (2017)
  • S. Moccia et al.

    Blood vessel segmentation algorithms — review of methods, datasets and evaluation metrics

    Comput. Methods Programs Biomed.

    (2018)
  • M.M. Fraz et al.

    Blood vessel segmentation methodologies in retinal images - a survey

    Comput. Methods Programs Biomed.

    (2012)
  • Kitware Inc. The visualization toolkit[Internet]. [cited 7Feb 2019]. Available:...
  • Kitware Inc. Insight segmentation and registration...
  • H. Johnson et al.

    The ITK Software Guide

    (2015)
  • L. Antiga

    Generalizing vesselness with respect to dimensionality and shape

    Insight J.

    (2007)
  • A.F. Frangi et al.

    Multiscale vessel enhancement filtering 1 introduction 2 method

    Med. Image Comput.

    (1998)
  • E.H.W. Meijering et al.

    Quantitative evaluation of convolution-based methods for medical image interpolation

    Med. Image Anal.

    (2001)
  • S. Bauer et al.

    A skull-stripping filter for ITK

    Insight J.

    (2012)
  • D. Mattes et al.

    PET-CT image registration in the chest using free-form deformations

    IEEE Trans. Med. Imaging

    (2003)
  • Y. Sato et al.

    Three-dimensional multi-scale line filter for segmentation and visualization of curvilinear structures in medical images

    Med. Image Anal.

    (1998)
  • J.L. Bentley

    Multidimensional binary search trees used for associative searching

    Commun. ACM

    (1975)
  • J.O. Neumann et al.

    Spatial distortion in MRI-Guided stereotactic procedures: evaluation in 1.5-, 3- and 7-Tesla MRI scanners

    Stereotact. Funct. Neurosurg.

    (2015)
  • L.R. Dice

    Measures of the amount of ecologic association between species

    Ecology

    (2006)
  • Cited by (0)

    View full text