Generic 3D Segmentation in Medicine based on a Self-learning

Topological Model

Gerald Zwettler

and Werner Backfrieder

1,2

Bio - and Medical Informatics, Research and Development Department, Upper Austria University of Applied Sciences,

Softwarepark 11, 4232 Hagenberg, Austria

School of Informatics, Communication and Media, Upper Austria University of Applied Sciences,

Softwarepark 11, 4232 Hagenberg, Austria

Keywords: Model-based Image Segmentation, Statistical Image Classification, Hybrid Watershed Pre-segmentation.

Abstract: Three-dimensional segmentation of medical image data is crucial in modern diagnostics and still subject of

intensive research efforts. Most fully automated methods, e.g. the segmentation of the hippocampus, are

highly specific for certain morphological regions and very sensitive to variations in input data, thus

robustness is not sufficient to achieve sufficient accuracy to serve in differential diagnosis. In this work a

processing pipeline for robust segmentation is presented. The flexibility of this novel generic segmentation

method is based on entirely parameter-free pre-segmentation. Therefore a hybrid modification of the

watershed algorithm is developed, employing both gradient and intensity metrics for the identification of

connected regions depending on similar properties. In a further optimization step the vast number of small

regions is condensed to anatomically meaningful structures by feature based classification. The core of the

classification process is a topographical model of the segmented body region, representing a sufficient

number of features from geometry and the texture domain. The model may learn from manual segmentation

by experts or from its own results. The novel method is demonstrated for the human brain, based on the

reference data set from brainweb. Results show high accuracy and the method proves to be robust. The

method is easily extensible to other body regions and the novel concept shows high potential to introduce

generic segmentation in the three-dimensional domain into a clinical work-flow.

1 INTRODUCTION

The accurate and preferably fully-automated

segmentation of medical image data is of high

importance for a broad range of medical

applications. The importance of computer-based

support for surgery planning, disease monitoring and

general diagnostics, by allowing for precise

estimation of volume, size and relative position of

anatomical structures, will constantly grow in

clinical practice. As an example, after automated

segmentation of liver parenchyma, hepatic vessels

and possible lesions utilizing level sets, the tumour

position can be analyzed with respect to the

supporting vessels and liver lobes which is of high

importance for surgery planning (Zwettler et al.,

2009). The informative value image data from the

functional imaging domain like SPECT or PET can

be raised by combining high anatomical resolution

of tomographic modalities like MRI and CT.

Thereby, the metabolic activity can be quantified

utilizing patient specific segmentation masks derived

from the anatomical imaging (Beyer et al., 2010).

In this work we present a concept for model-

based segmentation of 3D tomographic medical

image data based on a generic, parameter-free pre-

segmentation process and texture feature driven

region merging for classification. Our novel pre-

segmentation strategy combines aspects of gradient

based watershed transform (Beare and Lehmann,

2006), confidence connected region growing, region

merging and new variations in a hybrid algorithm

(Zwettler and Backfrieder, 2012). Starting at local

minima positions besides gradient height of classical

watershed transform, region intensity statistics are

used as merge metric. For region merging of the

initially pre-processed regions, several metrics,

namely watershed level tolerance, geometric

properties and similarity of the intensity profile are

combinded. Thus, an arbitrary input image can be

104

Zwettler G. and Backfrieder W..

Generic 3D Segmentation in Medicine based on a Self-learning Topological Model.

DOI: 10.5220/0004294701040108

In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2013), pages 104-108

ISBN: 978-989-8565-48-8

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

pre-classified at a user specified number of target

regions, defining the granularity of this pre-

processing step. Based on the pre-segmented

regions, the final segmentation is achieved by

feature-based classification utilizing cost

optimization. Besides texture features, also

geometric properties are incorporated, all derived

from a statistical a priori model calculated from

manual reference segmentations. For precise

reference segmentations at low user interactions,

rapid prototyping image processing chains have

been evaluated. The graph-based topographic

modelling of the anatomical context to segment

allows the segmentations at different hierarchical

levels. Applicability for future multi-modal image

processing will be evaluated in future.

2 DATA

For testing purposes concerning manual reference

segmentations, automated pre-segmentation and

feature-based classification, n=20 T1-weighted MRI

datasets from the simulated brainweb database

(Kwan et al., 1999) and the associated reference

segmentations are used.

Further test runs and validations are performed

utilizing n=12 anonymous multi-modal patient

studies, comprising morphologic image acquisitions

(T1, T2, PD,...) as well as related functional imaging

(SPECT, PET). For the patient data sets, the required

reference segmentations required for model training

and leave-one-out validation are achieved in a semi-

automated way by applying image processing

pipelines utilizing MeVisLab modules (Ritter, 2007)

a discussed in the later sections.

3 METHODOLOGY

In the preparation phase of our segmentation

concept, model definition is performed with respect

to the imaging modality to support and the

hierarchical anatomical topography of interest.

Furthermore, the classification features are chosen

with respect to their correlation. Based on the

anatomical topography and the chosen features, a

sufficient set of reference segmentations has to be

processed applying a semi-automated image

processing chain to evaluate the model parameters

with respect to each particular feature and each

particular anatomical structure to consider.

After preparing and training the model,

pre-segmentation of the tomographic patient dataset

to process is performed, utilizing a hybrid approach

incorporating aspects of watershed transform,

confidence connected region growing and region

merging. The chosen features are evaluated for all

regions resulting from the pre-segmentation process

step. The final segmentation is interpreted as

optimization problem, as the anatomical structure

labels are assigned to the particular regions in a way

to achieve a classification result that minimizes the

overall error with respect to the statistical feature

values, see Fig 1.

Figure 1: Illustration of our model-based segmentation

strategy. After definition of anatomical topography and

model training as preparation, pre-segmentation of the 3D

tomographic data is performed and resulting regions are

classified according to statistical features for final

segmentation.

3.1 Definition of Hierarchical

Anatomical Topography

For each anatomical structure to segment in the

particular medical context and for the specified

imaging modality, the position within the cascading

hierarchy of granularity has to be defined, see Fig. 2.

At level 0 the first separation into foreground as the

region of interest and the background as remaining

voxels is performed. Later the structures are

subdivided into composing structures according to

anatomical topography.

Generic3DSegmentationinMedicinebasedonaSelf-learningTopologicalModel

105

Figure 2: Illustration of topographic modelling.

Granularity is increased at higher levels.

3.2 Hybrid Pre-Segmentation

A generic segmentation of arbitrary image data can

be achieved utilizing an image processing chain,

including anisotropic diffusion filtering for

smoothing image data, gradient magnitude

extraction and a hybrid watershed implementation,

incorporating intensity homogeneity besides

gradient borders and neighbour region boundary

conditions as explained in (Zwettler and

Backfrieder, 2012) and illustrated in Fig. 3.

Figure 3: The sequential process chain (solid red arrows)

comprises low-pass filtering, high-pass filtering, minima

detection region growing, the watershed-type

segmentation procedure and finally region merging until

the convergence criterion is reached. Filtered input image

and the gradient magnitude representation are required as

input for several particular process steps (dashed green

arrows).

3.3 Feature-based Classification

Our segmentation strategy is defined as hierarchical

multi-feature optimization problem for classification

of pre-defined anatomical structures with respect to

their statistical feature values. For each anatomical

object at each hierarchy level, a statistical feature

vector must be preserved as a-priori information.

The chosen features comprise classic metrics

derived from co-occurrence matrix, like entropy and

energy (Felipe, 2003), the deviation of the intensity

profile, results of structural analysis with Hessian

eigenvalues (Sato et al., 1998) and others. Beside

these textural parameters, shape information, like

deviation of the size and extent is incorporated.

The segmentation itself is now defined as an

optimization problem of assigning the pre-classified

3D elements to the defined objects at lowest

hierarchy level to minimize the cumulated total

feature error. Due to recursive dependencies, only

one hierarchy level is optimized at each time and the

results are used as input for the next level in an

iterative top-down and bottom-up cycle. As first

initialization, the best matching regions for the

particular anatomical structures are iteratively

assigned with respect to expected size of the

particular structure. Results after this first

classification run are currently our final results.

In future, refinement of the segmentation results

is expected to be achieved utilizing heuristic

optimization based on evolution strategy

(Rechenberg, 1973) and classic genetic algorithms

(Goldberg, 1989).

4 RESULTS

Utilizing hybrid pre-segmentation, precision of

common watershed segmentation can be increased

from 0.88 to 0.91 and 0.92 respectively for our static

and adaptive intensity interval region merging on

average. The target number of regions to be pre-

segmented is reliably reached and misclassifications

only occur in border areas between neighbouring

anatomical structures, see Fig. 4.

Figure 4: Misclassified voxel (red) of AII4_WS

segmentation strategy in slices #70 and #150 compared to

brainweb reference segmentations displayed with respect

to original image data of dataset subject 04.

VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications

106

Concerning classification, for each feature the

prediction reliability is calculated based on

distributions with mean μ and variance δ calculated

as statistical average of all reference segmentations

and each particular anatomical structure to

distinguish. The reliability thereby indirectly

correlates with the error due to overlapping areas.

The overall prediction reliability is calculated via

stepwise integration over the entire feature value

range. The distributions of the anatomical structures

are normalized with respect to their overall

probability, i.e. statistical differences in anatomical

structure size.

Currently we evaluate 34 different feature types

for their prediction reliability when training an

anatomical topography model. Some of the most

notable are presented in Table 1 for analysis on

brainweb datasets. For construction of the feature

vector, a number of 10 of the best discriminating

features, should be selected. Despite choosing the

features with highest prediction reliability, it has to

be assured, that correlation within the feature vector

is low.

Table 1: Different types of features and their reliability for

classifying the different anatomical structure feature value

distributions.

feature

description prec.

1 maximum intensity value 67.93

3 median intensity value 87.07

4 mean intensity value 88.43

5 quantile 25 intensity 90.37

7 anatomical structure size 82.34

21 surface-to-volume-ratio 82.29

22 entropy of intensities 75.40

23 energy of intensities 74.12

25 mean probability of intensities 78.46

5 CONCLUSIONS

A generic segmentation concept for fast model-

based adjustment to particular image segmentation

tasks and imaging modalities has been presented.

Hybrid pre-segmentation is perfectly applicable for

context-free pre-processing of arbitrary image data

for first region labelling. Correlation of the analyzed

texture and geometric features shows promising

results for future heuristics-based classification

according to pre-defined anatomical topography.

The discussed and continuously refined rapid

prototyping image processing chain is perfectly

applicable for fast and robust preparation of

reference segmentations for training the a priori

model.

ACKNOWLEDGEMENTS

Thanks to our medical partners from the Wagner-

Jauregg state mental hospital, Linz, Upper Austria,

at the institute for neuro-nuclear medicine headed by

Primarius Dr.Dr. Robert Pichler for providing

medical image data and for valuable discussion.

This research is part of the INVERSIA project

(http://inversia.fh-linz.at) which was funded by the

European Regional Development Fund (ERDF) in

cooperation with the Upper Austria state

government (Regio13).

REFERENCES

Beare, R., and Lehmann, G., 2006. The watershed

transform in ITK – discussion and new developments.

In Insight Journal.

Beyer, T., Schwenzer, N., Bisdas, S., Claussen C.D., and

Pichler, B.J., 2010. MR/PET – Hybrid Imaging for the

Next Decade. In MAGNETOM Flash 3/2010.

Felipe, J. C., 2003. Retrieval by content of medical images

using texture for tissue identification. In CBMS.

Goldberg, D. E., 1989. Genetic Algorithms in Search

Optimization and Machine Learning. In Addison-

Wesley Professional.

Kwan, R. K.-S., Evans, A. C., Pike, G.B., 1999. MRI

simulation-based evaluation of image-processing and

classification methods. In IEEE Transactions on

Medical Imaging 18(11):1085-1097.

Rechenberg, I., 1973. Evolutionsstrategie-Optimierung

technischer Systeme nach Prinzipien der biologischen

Evolution. In Frommann-Holzboog Verlag, Stuttgart,

Germany.

Ritter, F., 2007. Visual Programming for Prototyping of

Medical Applications. In IEEE Visualization

workshop.

Sato, Y. S., Atsumi, H., Koller, T., Gerig, G., Yoshida, S.,

Kikinis, R., 1998. Three-dimensional multi-scale line

filter for segmentation and visualization curvilinear

structures in medical images. In Medical Image

Analysis 2 (2), 143-168.

Zwettler, G., Backfrieder, W., Swoboda, R., Pfeifer, F.,

2009. Fast Fully-automated Model-driven Liver

Segmentation Utilizing Slicewise Applied Levelsets

on Large CT Data. In Proc. of the 21st European

Modeling and Simulation Symposium, 161-166.

Generic3DSegmentationinMedicinebasedonaSelf-learningTopologicalModel

107

Zwettler, G., Backfrieder, W., 2012. A New Hybrid

Algorithm Based on Watershed Method, Confidence

Connected Thresholding and Region Merging as

Preprocessing for Statistical Classification of General

Medical Images. In Proc. of the 24th European

Modeling and Simulation Symposium.

VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications

108