Joint segmentation of bones and muscles using an intensity and histogram-based energy minimization approach

https://doi.org/10.1016/j.cmpb.2017.12.027Get rights and content

Highlights

  • A new energy minimization based approach for segmenting muscles and bones is proposed.

  • The proposed algorithm can be applied to surgery planning, disease diagnosis, analysis of fractures and/or bone/muscle densities.

  • The proposed energy function includes distance to histogram models of bone/muscle combined with gray-level information.

  • The algorithm outperforms other state-of-the art multi-label segmentation schemes.

Abstract

Background and objectives

The segmentation of muscle and bone structures in CT is of interest to physicians and surgeons for surgical planning, disease diagnosis and/or the analysis of fractures or bone/muscle densities. Recently, the issue has been addressed in many research works. However, most studies have focused on only one of the two tissues and on the segmentation of one particular bone or muscle. This work addresses the segmentation of muscle and bone structures in 3D CT volumes.

Methods

The proposed bone and muscle segmentation algorithm is based on a three-label convex relaxation approach. The main novelty is that the proposed energy function to be minimized includes distance to histogram models of bone and muscle structures combined with gray-level information.

Results

27 CT volumes corresponding to different sections from 20 different patients were manually segmented and used as ground-truth for training and evaluation purposes. Different metrics (Dice index, Jaccard index, Sensitivity, Specificity, Positive Predictive Value, accuracy and computational cost) were computed and compared with those used in some state-of-the art algorithms. The proposed algorithm outperformed the other methods, obtaining a Dice coefficient of 0.88 ± 0.14, a Jaccard index of 0.80 ± 0.19, a Sensitivity of 0.94 ± 0.15 and a Specificity of 0.95 ± 0.04 for bone segmentation, and 0.78 ± 0.12, 0.65 ± 0.16, 0.94 ± 0.04 and 0.95 ± 0.04 for muscle tissue.

Conclusions

A fast, generalized method has been presented for segmenting muscle and bone structures in 3D CT volumes using a multilabel continuous convex relaxation approach. The results obtained show that the proposed algorithm outperforms some state-of-the art methods. The algorithm will help physicians and surgeons in surgical planning, disease diagnosis and/or the analysis of fractures or bone/muscle densities.

Introduction

There is a growing tendency to use surgical planning tools prior to surgical interventions. Several surgical simulation software tools exist which either require complex/radiological workstations or have been designed for very specific types of surgery [1], [2], [3]. Others, such as MIMICS [4], [5] and AYRA [6], are more general in scope and have been used in almost all kinds of surgery. AYRA is the trade name of VirSSPA©, a software system which has been approved by medical regulatory bodies for routine clinical use. VirSSPA was developed as part of a research project conducted by a multidisciplinary team made up of researchers from the Virgen del Rocío University Hospital and the University of Seville (Spain) [7], [8], [9], [10]. To date, more than 1700 real cases have been analyzed using this tool.

However, the segmentation techniques implemented in AYRA encounter difficulties when attempting to segment bones and muscle accurately. Firstly, Hounsfield values [11] vary within these structures. Bones comprise different components, the principal ones being periosteum, compact bone, cancellous bone and bone marrow. Each of these has different Hounsfield values. There are also three different types of muscle tissue (skeletal, cardiac and smooth), with no established pattern common to all parts of the body. Moreover, Hounsfield values within these tissue types vary according to many factors, such as the physical condition of the person under study. Secondly, some diseases and radiation treatments modify the densities of bones and muscles [12], [13]. And thirdly, the 3D musculoskeletal system is geometrically complex and highly variable in shape.

The segmentation of bone and muscle structures has been widely addressed and it has been carried out using many different strategies, which could be classified into: intensity-based [12], [14], [15], edge-based [16], region-based [17] and model-based [18], [19], [20]. Some works combining these techniques have been used with certain success [14], [16].

Many of the studies in the literature focus on segmenting one particular type of bone [14], [21], [22], [23], [24], [25], [26] or muscle [18], [19], so it would be desirable to have a generalized, automatic method for segmenting a wider variety of bone and muscle structures.

The VirSSPA commercial surgical planning tool includes three segmentation algorithms for: a) thresholding; b) region growing; and c) adaptive region growing [17], [27]. Surgeons from the Virgen del Rocío hospital, in Seville, have reported difficulties when segmenting bones and muscles with these algorithms due to the reasons described above. In this paper, therefore, we propose a new method that will avoid these problems.

Image segmentation is today successfully carried out using model based techniques [20], [21], [22], [23], [24], [25], [26], [28], [29], [30], [31], [32], [33], [34], [35]. In this type of methodology energy measure, which includes region and boundary information, is minimized to solve the segmentation problem. One such algorithm developed over the last few years is that of minimization based on convex relaxation [36], [37], [38]. This solution has been used successfully to solve segmentation problems in different medical applications [39], [40], [41].

This paper presents an algorithm based on convex relaxation, designed to segment bone and muscle structures. The main novelty of this approach is that gray information and distance to histogram models of bones and muscles are combined to generate the cost functions that will subsequently be used for minimization.

Section snippets

Material

The dataset used in the experiments comprised 27 CT volumes corresponding to different parts of the body from 20 patients. All CT volumes were manually segmented in consensus by two experts. The manually segmented CTs were used as ground truth for evaluation purposes. The validation data and the ground truth (USevillabonemuscle database) have been made publicly available at http://grupo.us.es/grupobip/research/research-topics/segmentation-of-abdominal-organs-and-tumors/. The patients were aged

Results

To assess the algorithm's performance, a 3-fold cross-validation was performed [42]. At each iteration, the corresponding training set was used to obtain the Hounsfield ranges and histogram models for bone and muscle. Comparisons with some state-of-the-art methods were also carried out. The manual segmentations by the two experts were used as ground-truth. The techniques used for comparison purposes were:

  • 1.

    Thresholding. This technique is the one usually preferred by clinicians for segmenting

Discussion

This study addressed the general segmentation of skeletal muscle and bone structures. The objective was not easy and certain limitations were encountered during the course of the research. One such limitation was the resolution of the image acquisition device employed. In our approach, this problem was exacerbated due to the size of the local window we used to compute the 3D local histograms around the voxels under analysis. Device resolution and the size of the local windows used in the

Conclusions

This work has described an algorithm for segmenting muscle and bone structures using a continuous multi-label convex relaxation approach in which statistical information and gray level information are combined in the cost terms.

The algorithm has been compared with other state-of-the art continuous multi-label schemes. The results obtained show that the proposed algorithm outperforms the other techniques in all metrics except computational time, where the threshold algorithm clearly demonstrates

Author contributions

The proposed algorithm was developed by J.A Pérez-Carrasco, C. Serrano, C. Suárez-Mejías and B. Acha. The same authors were involved, together with J. L. López, in the validation of the algorithm. Moreover, all the aforementioned authors drafted, critically revised and approved the final version of the full manuscript, and have therefore contributed sufficiently to the scientific work carried out to share collective responsibility and accountability for the results thereof.

Conflict of interest statement

The authors declare that they have no competing interests.

Compliance with ethical standard

All procedures performed as part of the research were approved by the local Ethics Committee at the Virgen del Rocío University Hospital in Seville, Spain. The studies carried out were in compliance with the 1964 Declaration of Helsinki and its later amendments, or with comparable ethical standards.

In this research none of the authors carried out any experiments with animals.

Informed consent was obtained from all individual participants included in the study.

Acknowledgments

This research was co-financed by P11-TIC-7727, (Regional Government of Andalusia, Spain) and DPI2016-81103-R (Spanish Government State Research Plan Retos, 2013–2016).

References (58)

  • Proplan CMF software, June 2017. Accessed 11...
  • T. Clijmans et al.

    Computer supported pre-operative planning of craniosynostosis surgery: a MIMICS-integrated approach

  • MIMICS, June 2017. Accessed 11...
  • AYRA, June 2017. Accessed 11...
  • C. Suárez et al.

    VirSSPA- a virtual reality tool for surgical planning workflow

    Int. J. Comput. Assist. Radiol. Surg.

    (2009)
  • P. Gacto et al.

    A three-dimensional virtual reality model for limb reconstruction in burned patients

    BURNS

    (2008)
  • T. Gómez-Cía et al.

    The virtual reality tool VirSSPA in planning DIEP microsurgical breast reconstruction

    Int. J. Comput. Assist. Radiol. Surg.

    (2009)
  • P. Gacto-Sánchez et al.

    Computerised tomography angiography with VirSSPA 3D software for perforator navigation improves perioperative outcomes in DIEP flap breast reconstruction

    Plast. Reconstr. Surg.

    (2010)
  • T.G. Feeman

    The mathematics of medical imaging: a beginner's guide

    Springer Undergraduate Texts in Mathematics and Technology

    (2010)
  • Þ. Pétursson et al.

    Bone and muscle assessment in patients undergoing total hip arthroplasty using HU based analysis

    Eur. J. Translat. Myol.

    (2012)
  • D.R. Taaffe et al.

    Race and sex effects on the association between muscle strength, soft tissue, and bone mineral density in healthy elders: the health, aging, and body composition study

    J. Bone Mineral Res.

    (2001)
  • S. Banik et al.

    Automatic segmentation of the ribs, the vertebral column, and the spinal canal in pediatric computed tomographic images

    J. Digit. Imaging

    (2010)
  • Y. Kang

    A new accurate and precise 3-D segmentation method for skeletal structures in volumetric CT data

    IEEE Trans. Med. Imaging

    (2003)
  • C Mendoza et al.

    Self-Assessed Contrast-Maximizing Adaptive Region Growing

    (2009)
  • X.M. Pardo

    A snake for CT image segmentation integrating region and edge information

  • R.M. Rangayyan et al.

    Automatic delineation of the diaphragm in computed tomographic images

  • T. Shen et al.

    Active volume models for medical image segmentation

    IEEE Trans. Med. Imaging

    (2011)
  • J. Krátký et al.

    Three-dimensional segmentation of bones from CT and MRI using fast level sets

  • H. Li

    Contour extraction of hand-wrist skeletal based on active contour model

  • Cited by (17)

    • Patch-shuffle-based semi-supervised segmentation of bone computed tomography via consistent learning

      2023, Biomedical Signal Processing and Control
      Citation Excerpt :

      The first row was a right-foot example, the second row was a left-foot example, and the third row was a two-feet example. The USEvillaBone dataset [32] contained 27 CT scans, and each scan contained 10 slices. In total, there were 270 slices in the USEvillaBone dataset.

    • Bone segmentation on whole-body CT using convolutional neural network with novel data augmentation techniques

      2020, Computers in Biology and Medicine
      Citation Excerpt :

      Ground truth segmentation was performed by a medical expert and verified by a radiologist (6-years' experience in diagnostic imaging) for each scan. To compare our model with those of previous studies, we trained and validated our network on a publicly available labelled dataset [12]. The dataset comprised 270 slices of CT images from 27 scans, obtained in 20 patients.

    • Machine Learning for Automatic Paraspinous Muscle Area and Attenuation Measures on Low-Dose Chest CT Scans

      2019, Academic Radiology
      Citation Excerpt :

      Despite the prognostic value of CT measurements, the current approaches to measuring muscle CSA and MA require laborious manual or semiautomated tissue segmentation, which is not practical for large clinical trials or clinical practice (37–39). Various approaches have been used to automate muscle segmentation, with atlas-based (40–44) and convolutional neural network (CNN)-based (44–48) approaches showing the greatest promise. Popuri et al. (40) developed an atlas-based approach for automatic total abdominal muscle segmentation, reporting a similarity coefficient of 0.91 in a test set of 1069 abdominal CT images.

    • Robust semi-automatic segmentation method: an expert assistant tool for muscles in CT and MR data

      2024, Computer Methods in Biomechanics and Biomedical Engineering: Imaging and Visualization
    View all citing articles on Scopus
    View full text