gms | German Medical Science

GMDS 2012: 57. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie e. V. (GMDS)

Deutsche Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie

16. - 20.09.2012, Braunschweig

Influence of combination strategies and registration parameters on multi-atlas-segmentation of lungs and lung lobes

Meeting Abstract

Search Medline for

  • Christiane Steinberg - Universität zu Lübeck, Deutschland
  • René Werner - Universität zu Lübeck, Deutschland
  • Heinz Handels - Universität zu Lübeck, Deutschland

GMDS 2012. 57. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie e.V. (GMDS). Braunschweig, 16.-20.09.2012. Düsseldorf: German Medical Science GMS Publishing House; 2012. Doc12gmds051

DOI: 10.3205/12gmds051, URN: urn:nbn:de:0183-12gmds0517

Published: September 13, 2012

© 2012 Steinberg et al.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License ( You are free: to Share – to copy, distribute and transmit the work, provided the original author and source are credited.



Introduction: Segmentation of lungs and lung lobes is a prerequisite for pulmonary image analysis/diagnostics (e.g. detection of abnormalities in the lungs) [1]. It therefore attracts a great interest (cf. the LOLA11 workshop,, but in clinical practice, segmentation is often performed at least partly manually, which is error prone and time consuming; automatic segmentation would be clearly desirable. A promising approach would be multi-atlas-segmentation. This approach, however, requires performing a series of usually time consuming non-linear registration runs, which still can be critical in clinical environment [2]. Applying a common registration framework and considering different combination strategies for multi-atlas-segmentation of lung and lung lobes in chest CT data, in this paper the influence of registration parameters is studied, keeping especially the trade-off between registration speed and segmentation accuracy in mind.

Material and Methods: Data Sets: Nine chest CT scans (256x256x278 voxels) of patients without obvious lung pathologies were used. Segmentations were generated manually and served as ground truth. For evaluation purposes, a leave-one-out cross validation has been performed (segmentation of a single chest CT by using the eight remaining atlases) for each combination of the parameters described subsequently.

Registration: As a first approach, atlas images and the CT to be segmented were registered using an iterative non-linear intensity-based non-parametric diffusion approach, implemented in a multi-resolution framework (here: three levels). As the registration on the finest level (full image resolution) is most time-consuming, we varied the number of iterations on this level (between 1 and 500). Further, non-linear registration was combined with an affine pre-registration. The results were compared with a segmentation based on only an affine image alignment. Combination strategies: For all registration settings, the combination strategies vote and sum rule [3] were applied and compared. Results were additionally compared with a single-atlas segmentation using the "best atlas after affine" and the "best atlas after non-linear registration" of the atlases and the CT to be segmented (measure: SSD).

Results: For our test data sets, no significant differences were observed between the combination strategies considered (p>0.05, t-test of Dice overlap coefficients). Mean Dice values after only affine registration were 0.87 for the whole lungs and 0.78 for lung lobes (mean runtime: several seconds); segmentation based on only non-linear registration failed for some cases. When combining affine and non-linear registration, the Dice values increased to 0.99 (lungs) and 0.89 (lobes) after 500 iterations on the finest level. Interestingly, reducing the number of iterations to only one did not change the Dice values – with runtimes of a single CT-CT non-linear registration being 3 instead of 20 minutes.

Discussion: The results show that already a few iterations (here even a single one) on the finest and most time consuming registration level lead to adequate segmentation results. Thus, a key point of criticism of multi-atlas-segmentation methods (high expenditure of time) can be invalidated for the given application. Of course, the presented study is based on only a limited number of patients and the results have to be verified using a larger data set.


Sluimer I, Schilham A, Prokop M, van Ginneken B. Computer analysis of computed tomography scans of the lung: A survey. IEEE Transactions on Medical Imaging. 2006;25(4):385-405.
van Rikxoort EM, Isgum I, Staring M, Klein S, van Ginneken B. Adaptive Local Multi-Atlas Segmentation: Application to Heart Segmentation in Chest CT Scans. Proceedings of SPIE. 2008;6914:691407.
Kittler J, Alkoot FM. Sum versus Vote Fusion in Multiple Classifier Systems. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2003;25(1):110-5.