Elsevier

Medical Image Analysis

Volume 8, Issue 3, September 2004, Pages 245-254
Medical Image Analysis

Automated segmentation of the left ventricle in cardiac MRI

https://doi.org/10.1016/j.media.2004.06.015Get rights and content

Abstract

We present a fully automated deformable model technique for myocardium segmentation in 3D MRI. Loss of signal due to blood flow, partial volume effects and significant variation of surface grey value appearance make this a difficult problem. We integrate various sources of prior knowledge learned from annotated image data into a deformable model. Inter-individual shape variation is represented by a statistical point distribution model, and the spatial relationship of the epi- and endocardium is modeled by adapting two coupled triangular surface meshes. To robustly accommodate variation of grey value appearance around the myocardiac surface, a prior parametric spatially varying feature model is established by classification of grey value surface profiles. Quantitative validation of 121 3D MRI datasets in end-diastolic (end-systolic) phase demonstrates accuracy and robustness, with 2.45 mm (2.84 mm) mean deviation from manual segmentation.

Introduction

Cardiac MRI provides important information for diagnosis and treatment of cardiovascular diseases by enabling quantitative assessment of functional parameters such as wall motion, wall thickness and ejection fraction (Frangi et al., 2001). However, the lack of computational tools still implies that not all the information contained in the data is currently used. In order to fully exploit the information in the MRI data in clinical practice, and considering the amount of data involved (10–20 frames by 8–12 image slices), automated extraction of the myocardium is essential. However, poor object features resulting from signal loss due to blood flow and partial volume effects, and significant variation of grey value appearance make finding a robust and accurate solution a difficult problem.

Recent approaches to automated segmentation incorporate prior knowledge to improve robustness (Jolly et al., 2001; Lorenzo-Valdes et al., 2002; Mitchell et al., 2001; Montagnat and Delingette, 2000). Lorenzo-Valdes et al. (2002) achieved convincing results by volumetric atlas matching using B-spline registration. However, this approach requires several minutes for a single time frame. The fast and robust framework of active shape and appearance models by Cootes et al. (1998) was extended and applied to cardiac 3D and 2D time series (Mitchell et al., 2001), achieving promising results. However, statistical shape models cannot capture variability outside the learning set, which is likely to occur in the case of pathology, and appearance modeling may fail in the presence of large grey value variability across subjects and time. Learning sets could be used for specific targeted pathologies, but this is time consuming and requires a large database of pathologic cases.

Earlier work by this group proposed shape constrained deformable models, enabling local deviation from a statistical shape model by embedding it into a 3D triangular deformable mesh (Kaus et al., 2003; Weese et al., 2001). This approach used deterministic feature search and was successfully applied to segment bones in CT. However, like many other deformable model approaches (McInerney and Terzopoulos, 1996), multiple objects like, e.g., the endo- and epicardium are not considered, or objects of significant shape variation over time and across subjects. In addition, methods with deterministic feature functions assume constant grey value appearance of the object’s surface, which is not true for the myocardiac surface due to different surrounding tissue as, e.g., lung parenchyma, fat, and the blood pool in the right ventricle.

We propose to integrate several prior information sources into deformable models including a deterministic, parametric model of the variation of surface features, inter- and intra-subject shape variation, and spatial relationships to handle multiple objects. In the remainder of this article we present the modeling and segmentation methods with application to cardiac 3D MRI time series.

Section snippets

Methods

In this section we outline the deformable model framework (Weese et al., 2001) and then describe how to represent, learn and integrate prior knowledge into the method.

Experimental results

For 3D adaptation of the shape model, automatic initial positioning of the shape model in the center of the image volume was sufficient for the cases in our image database. In cardiac imaging, the location of the heart can be assumed approximately in the center of the image. To increase the capture range of our method, we adopt a coarse-to-fine adaptation approach on an image resolution pyramid generated with a fast B-spline based method (Lee et al., 1998). Because the in-plane resolution of

Conclusion

We presented a method for modeling and fully automated segmentation of the LV in 3D cardiac MRI time-series with a mean error of 2.45 ± 0.75 mm (2.84 ± 1.05 mm) for the end-diastolic (end-systolic) phase. The method is fast and requires approximately 7 s on standard hardware for the segmentation of a heart phase.

We demonstrated that the integration of multi-model prior knowledge into a deformable model is possible in an efficient way, and significantly increases the robustness and the accuracy of

Acknowledgements

Cardiac MRI have been provided by Warren J. Manning, MD, Co-Director of the Cardiac MR Center of the Beth Israel Deaconess Medical Center, Boston, MA. We thank Miriam Siers, Image Sciences Institute, University Medical Center, for the manual segmentation of the cardiac MRI. We thank our colleagues C. Lorenz, T. Netsch, K. Meetz, Philips Research Hamburg, and S. Lobregt, R. Truyen, F. Gerritsen, Philips Medical Systems Best, for useful discussions. M.R.K. thanks D. Rückert, Imperial College

References (21)

  • T. McInerney et al.

    Deformable models in medical image analysis: a survey

    Med. Image Anal.

    (1996)
  • J. Montagnat et al.

    Globally constrained deformable models for 3D object reconstruction

    Signal Process.

    (1998)
  • T. Cootes et al.

    Combining point distribution models with shape models based on finite element analysis

    Image Vision Comput.

    (1995)
  • T. Cootes et al.

    Active appearance models

  • R. Duda et al.

    Pattern Classification and Scene Analysis

    (1973)
  • P. Fletcher et al.

    Deformable m-rep segmentation of object complexes

  • A. Frangi et al.

    3D modeling for functional analysis of cardiac images: a review

    IEEE Trans. Med. Imaging

    (2001)
  • P. Gill et al.

    Practical Optimization

    (1981)
  • G. Golub et al.

    Matrix Computations

    (1996)
  • M.-P. Jolly et al.

    Segmentation of the left ventricle in cardiac MR images

There are more references available in the full text version of this article.

Cited by (231)

  • Left ventricle automatic segmentation in cardiac MRI using a combined CNN and U-net approach

    2020, Computerized Medical Imaging and Graphics
    Citation Excerpt :

    Inspired by this, many complexity methods have been proposed. Kaus M.R. presented a deformable model for the segmentation of LV in 3D cardiac MRI time-series (Kaus et al., 2004). Feng C. developed an LV segmentation method in CMR short-axis images using distance regularized two-layer level set approach (Feng et al., 2013).

View all citing articles on Scopus
View full text