Multi-organ 3D image segmentation

Problem and context

Today's clinical practice and medical research generate large numbers of high-resolution 3D images, which make tasks of efficient data access, transfer, analysis and visualization challenging, especially in today’s distributed computing environments. The segmentation of objects in images is often a precondition to their efficient analysis, manipulation and visualization. There is great interest in robust and fast segmentation algorithms for numerous applications such as the creation of anatomical patient-specific models, clinical applications, the creation of custom-made prosthetics, education, etc.

The segmentation of medical images presents a number of difficulties. Limitations of traditional segmentation approaches that we have tackled are:

  • Spatio-temporal algorithmic complexity linked to 3D images.
  • Simultaneous, as opposed to sequential, segmentation of all visible objects in the image.
  • Segmentation of objects having similar intensity profiles.
  • Modeling and accounting for inter-object spatial relationships in the algorithm.
  • Variability of tissues and organs across patient populations.

 

Methods, contributions and results

Semi-automatic multi-organ segmentation

Figure 1 An outline of our multi-object semi-automatic segmentation method applied to a synthetic image.

We have developed a generic Bayesian method for the efficient segmentation of 3D medical images which can be easily adapted to different applications [1, 2]. The segmentation is expressed as a Bayesian labeling problem (a label for each object), formulated in terms of a prior distribution described by a Markov random field, and a likelihood distribution derived from the image using “seed” voxels marked by an operator. The optimal segmentation, which corresponds to the maximum a posteriori solution (MAP), can be found efficiently by minimizing an energy function of unary and pairwise terms via multi-label Graph Cut algorithms [3], provided that pairwise terms are sub-modular. Our contributions are two-fold:

  1. A structural prior model, called the vicinity model. Sub-modular by definition, it incurs several levels of penalization in  multi-object  segmentation  and captures the spatial configuration of image objects. It is defined according to the shortest path metric on the adjacency graph of image objects.
  2. Reduction of memory budget and optimization runtime up to an order of magnitude without compromising the segmentation quality by clustering image voxels prior to segmentation by an adaptive centroidal Voronoi tessellation. We formulate Graph Cut segmentation according to the irregular graph of clusters such that its energy is independent of clustering resolution.

Qualitative and quantitative evaluations have confirmed the advantages of our vicinity prior model over the standard Potts model especially in the correct segmentation of objects having similar intensities, in the precise placement of object boundaries as well as in the robustness  of segmentation with respect to the clustering resolution. Figure 1 gives a top-level view of the approach and its original contributions and Figure 2 illustrates some of its results.

Figure 2 3D views of surface meshes extracted from structure/organ volumes segmented via our semi-automatic multi-object segmentation method.

 

Automatic multi-organ segmentation

During 2014-16, we have participated in Visceral benchmarks in collaboration with GIPSA-lab. We have extended our semi-automatic segmentation method by exploiting information on organ position, learned from the Visceral training dataset [4] by a fast image registration approach based on SURF keypoints. This approach allows to construct probabilistic atlases capturing organ position and shape which are subsequently introduced in the framework of Graph Cut segmentation. The image registration approach has also allowed to eliminate the need for “seed” voxels used to estimate intensity likelihoods, and consequently to render the algorithm completely automatic. These seeds are now deduced from image regions corresponding to high position probabilities in organ atlases which are registered to the image via a hierarchical application of the SURF-based image registration method [5].  Figure 3 gives a top-level view of the extended segmentation method applied to a CT image. The evaluation of our method as part of the third Visceral benchmark has shown that it ranks among the first two methods in the segmentation of each of the 20 tested organs over the 5 segmentation methods which have been evaluated.

Figure 3 A top-level view of our automatic multi-object segmentation method applied to a CT image.

 

References

  1. R. Kéchichian, S. Valette, M. Desvignes and R. Prost, “Shortest-Path Constraints for 3D Multiobject Semiautomatic Segmentation via Clustering and Graph Cut,” IEEE Trans. on Image Processing, vol. 22, no. 11, pp. 4224-4236, 2013.
  2. R. Kéchichian, “Structural Priors for Multiobject Semiautomatic Segmentation of Three-dimensional Medical Images via Clustering and Graph Cut Algorithms,”  PhD thesis, INSA-Lyon, Université de Lyon, 2013.
  3. Y. Boykov, O. Veksler, and R. Zabih, “Fast approximate energy minimization via graph cuts,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 23, no. 11, pp. 1222–1239, 2001.
  4. G. Langs, H. Müller, B. Menze and A. Hanbury, VISCERAL: towards large data in medical imaging - challenges and directions. Proc. MICCAI 2012 Workshop on Medical Content-based Retrieval for Clinical Decision Support (MCBR-CDS), 2012.
  5. R. Kéchichian, S. Valette, M. Sdika and M. Desvignes, “Automatic 3D multiorgan segmentation via clustering and graph cut using spatial relations and hierarchically-registered atlases,” MICCAI Medical Computer Vision: Algorithms for Big Data, 2014, pp. 201-209.