Elsevier

Medical Image Analysis

Volume 18, Issue 3, April 2014, Pages 567-578
Medical Image Analysis

Focused shape models for hip joint segmentation in 3D magnetic resonance images

https://doi.org/10.1016/j.media.2014.02.002Get rights and content

Highlights

  • We introduce a weighted shape learning approach applied to the human hip joint.

  • Weights may be set to focus the shape representation energy to important areas.

  • The highly weighted areas become the dominant representations within the model.

  • Lower reconstruction errors and higher accuracy can be obtained in those areas.

  • Validation was done on 35 3T unilateral small field of view MR scans.

Abstract

Deformable models incorporating shape priors have proved to be a successful approach in segmenting anatomical regions and specific structures in medical images. This paper introduces weighted shape priors for deformable models in the context of 3D magnetic resonance (MR) image segmentation of the bony elements of the human hip joint. The fully automated approach allows the focusing of the shape model energy to a priori selected anatomical structures or regions of clinical interest by preferentially ordering the shape representation (or eigen-modes) within this type of model to the highly weighted areas. This focused shape model improves accuracy of the shape constraints in those regions compared to standard approaches. The proposed method achieved femoral head and acetabular bone segmentation mean absolute surface distance errors of 0.55±0.18mm and 0.75±0.20mm respectively in 35 3D unilateral MR datasets from 25 subjects acquired at 3T with different limited field of views for individual bony components of the hip joint.

Introduction

Active Shape Models (ASMs) developed by Cootes et al. (1995) (also referred to as deformable models) have been successfully applied to segmenting three dimensional (3D) MR images of musculoskeletal structures (Neubert et al., 2012, Schmid et al., 2011, Fripp et al., 2010) as well as for other structures such as the prostate (Martin et al., 2010, Chandra et al., 2012), liver (Heimann et al., 2006) and the heart (Zheng et al., 2008, Ecabert et al., 2008). In this approach, a triangulated surface is deformed to fit an object of interest while simultaneously utilising other image features and/or priors, such as the object shape and appearance. For example, Fripp et al. (2010) used deformable models to (fully) automatically segment the articulating bone elements of the knee joint (femur, tibia and patella) from 3D T2-weighted water-excited Double-Echo Steady State (weDESS) MR images utilising shape priors for each bone. The bone segmentations were then used to accurately localise the bone-cartilage interfaces (BCIs) from which individual joint cartilage plates can be segmented and cartilage thickness determined for early Osteoarthritis (OA) assessment (Altman et al., 2004).

Accurate 3D segmentations of the osseous structures of the hip joint are important as current OA assessments utilise plain radiographs that are inconsistent in the early disease stages as they rely on bony features (abnormalities) that typically occur in later pathological stages of OA (Pollard et al., 2008). Most of the previous bone segmentation work has been performed using computerised tomography (CT) (Lamecker et al., 2004, Seim et al., 2008, Audenaert et al., 2011, Masjedi et al., 2012), plain radiographs (Ding et al., 2007, Dong et al., 2007, Zheng et al., 2010, Baka et al., 2011) or ultrasound images (Barratt et al., 2008). MR imaging enables noninvasive, 3D assessment of the joint structure including biochemical changes with no ionizing radiation and excellent soft tissue contrast (Burstein et al., 2000). This enables the simultaneous visualisation of both the bone and cartilage tissue as opposed to other imaging modalities. Segmentation of these osseocartilaginous structures of the hip joint is however more challenging in 3D MR images due to the presence of partial bone coverage resulting from the large joint geometry, relatively thin cartilage plates that are present (Hodler et al., 1992) and the partial voluming errors caused by its spherical structure (Naish et al., 2006).

Schmid et al. (2011) developed a robust shape model (RSM) approach incorporating the robust Principal Component Analysis (PCA) of Skočaj et al. (2007) to segment pelvic and femoral bone elements, visualised as partial structures, in small field of view (FoV) MR images. Whilst they reported encouraging results for bone segmentation with an average distance error of 1.12±0.46mm, further improvements in these bone segmentations, especially for both the femoral and acetabular BCI regions, would likely be of benefit for subsequent analyses of the relatively thin (1–2 mm thick) hip cartilages (Shepherd and Seedhom, 1999). This improved accuracy would also be particularly important in subsequent segmentation of the separated cartilage plates. Other techniques for segmenting bones, such as utilising the image information in the scans directly, include the works of Dalvi et al., 2007, Zoroofi et al., 2004, Nguyen et al., 2007, Bourgeat et al., 2007.

In this work, we introduce a weighted shape learning approach (via the spatially weighted PCA of Thomaz et al. (2010)) for region-specific accuracy and compactness applied to the human hip joint (see Fig. 1). Predetermined weights may be set to each corresponding point of an anatomically selected region in the shape model allowing the focusing of the shape constraint energy on areas deemed most important from a research or clinical perspective. This focused shape model (FSM) orders the shape representation to the up-weighted regions, in contrast to other approaches, where the representation is ordered from the largest variations to the smallest, regardless of the regions that might be of more importance clinically. In the FSM, the variation in the highly weighted areas become the dominant representations within the model and the resulting eigen-modes are ordered by importance to the problem at hand.

The predicted outcomes from our work with the FSM approach are that it provides:

  • 1.

    A more compact shape representation which requires substantially fewer modes to achieve equivalent accuracy to the RSM proposed by Schmid et al. (2011) for small FoV MR images when focusing on the hip joint.

  • 2.

    Lower reconstruction errors and higher accuracy when using the same precision of shape representation for the hip joint when compared to the RSM.

The paper is structured as follows. In the next section, the background information for this work is covered. Section 3 introduces the proposed FSM and Sections 4 Results, 5 Discussion present the results and discussion for its application to the hip joint respectively. The experiments conducted compare the FSM to the work of Schmid et al. (2011) on small FoV MR image bone segmentation with the eventual goal of utilising these as the basis for segmenting the femoral head and acetabular cartilages of the hip joint in future work.

Section snippets

Background

In medical imaging, it is important to keep the shape of the object of interest diagnostically interpretable so that it can be useful to a clinician. Typically at the deformation stage of the ASM, a surface S (having vertices vi with 0i<N, where N is the total number of vertices) is deformed in a direction that is normal to the surface at each vertex, limited only by an allowed maximum length of displacement m and independently of its neighbouring vertices. Neighbouring vertices can be

Focused shape models

In a number of applications it is necessary to get a more accurate segmentation of a specific region (s) of an object. For example, our eventual goal is to segment the hip cartilages for which some methods rely on an accurate bone segmentation near the BCI as priors (Fripp et al., 2010). By utilising the spatially weighted PCA, the BCI area of the hip joint can be up-weighted to account for important areas, all the while utilising the other areas to ensure proper pose alignment and more

Results

To assess the performance of the FSMs, a comparison was made to an in-house implementation of the RSM proposed by Schmid et al. (2011) (with one minor improvement) in three experiments with respect to the hip joint (noting that the eventual aim is to segment the hip joint cartilages in future work). Firstly, the generalisability (i.e. the reconstruction errors of the training surfaces using a leave-one-out approach) and relative compactness (i.e. the dimensional reduction of the shape

Discussion

The FSMs were found to have a highly compact shape representation that typically required only a few modes to achieve the equivalent accuracy as the RSM for segmentation of the bony components of the hip joint in small FoV MR images. Alternatively, lower reconstruction errors and higher accuracy were obtained for bone segmentations when using the same precision of shape representation as the RSM. This improvement in the BCIs of the hip joint (around 0.2–0.4 mm on average), most notably for the

Conclusion

This paper presented a weighted shape learning approach for deformable models applied to hip joint segmentation in 3D MR images. A spatially weighted PCA was used to build a focused shape model (FSM) that allowed selective up-weighting of the hip joint area to create a region-sensitive and region-compact shape representation for the joint. The up-weighting also resulted in an ordering of the model primary modes that was more related to the local anatomy than using standard approaches. The

Acknowledgments

This research was supported under Australian Research Council’s Linkage Projects funding scheme LP100200422 and partially funded by the Cancer Council NSW (Project Grants 07-06 and 11-05).

References (58)

  • T. Nishii et al.

    Three-dimensional distribution of acetabular cartilage thickness in patients with hip dysplasia: a fully automated computational analysis of MR imaging

    Osteoarthr. Cartilage

    (2004)
  • J. Schmid et al.

    Robust statistical shape models for MRI bone segmentation in presence of small field of view

    Med. Image Anal.

    (2011)
  • K.-k. Shen et al.

    Detecting global and local hippocampal shape changes in Alzheimer’s disease using statistical shape models

    NeuroImage

    (2012)
  • D. Skočaj et al.

    Weighted and robust learning of subspace representations

    Pattern Recognit.

    (2007)
  • C.E. Thomaz et al.

    A new ranking method for principal components analysis and its application to face image analysis

    Image Vision Comput.

    (2010)
  • T. Vercauteren et al.

    Diffeomorphic demons: efficient non-parametric image registration

    NeuroImage

    (2009)
  • P.A. Yushkevich et al.

    User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability

    NeuroImage

    (2006)
  • R.A. Zoroofi et al.

    Automated segmentation of necrotic femoral head from 3D MR data

    Comput. Med. Imaging Graphics

    (2004)
  • E.A. Audenaert et al.

    Development of a three-dimensional detection method of cam deformities in femoroacetabular impingement

    Skeletal Radiol.

    (2011)
  • P.J. Besl et al.

    A method for registration of 3-D shapes

    IEEE Trans. Pattern Anal. Mach. Intell.

    (1992)
  • D. Burstein et al.

    MRI techniques in early stages of cartilage disease

    Invest. Radiol.

    (2000)
  • S. Chandra et al.

    Patient specific prostate segmentation in 3-D magnetic resonance images

    IEEE Trans. Med. Imaging

    (2012)
  • Dalvi, R., Abugharbieh, R., Wilson, D., Wilson, D.R., 2007. Multi-contrast MR for enhanced bone imaging and...
  • M. Datar et al.

    Geometric correspondence for ensembles of nonregular shapes

    Med. Image Comput. Comput.–Assist. Intervention – MICCAI

    (2011)
  • R.H. Davies et al.

    Building 3-D statistical shape models by direct optimization

    IEEE Trans. Med. Imaging

    (2010)
  • L.R. Dice

    Measures of the amount of ecologic association between species

    Ecology

    (1945)
  • F. Ding et al.

    Automatic segmentation of femur bones in anterior-posterior pelvis X-ray images

  • P. Dodin et al.

    A fully automated system for quantification of knee bone marrow lesions using MRI and the osteoarthritis initiative cohort

    J. Biomed. Graphics Comput.

    (2012)
  • X. Dong et al.

    Automatic extraction of femur contours from calibrated x-ray images using statistical information

    J. Multimedia

    (2007)
  • Cited by (0)

    1

    Principal corresponding author.

    View full text