Automatic Determining of Vertebrae from CT Images

B. A. Zalesky, A. M. Nedzved, S. V. Ablameiko and P. V. Lukashevich

United Institute of Informatics Problems of the National Academy of Sciences of Belarus

Surganova 6, Minsk, Belarus

Abstract. Position of the spinal column and vertebrae on the CT images is one

of the main features to determine position of a patient and his organs. In this

paper we propose the algorithm to extract the spinal column and vertebrae with

ribs from CT images and to estimate their coordinates as well as coordinates of

human body in the coordinate system of a CT scanner.

1 Introduction

In recent decades remarkable improvements in diagnostic managements of oncology

patients have been achieved. We became witnesses of a technological breakthrough

in the field of diagnostic imaging, including the computerized tomography (CT),

magnetic resonance imaging (MRI), angiography etc.

Over the past decade, the increasing technological sophistication of CT equipment

has permitted excellent images with a limited dose of radiation which is currently

equal to a plain chest X-ray (0,001 Gy). Modern CT scanners were equipped with

special programs in order to save the information in DICOM files, to analyze images

and to provide interactive measurements. The programs disclose internal object struc-

tures and compute a great number of organ parameters. They make easier interactive

measurements of organ size and volume and identification of the character of the

pathological mass.

At present there are a few special systems and complexes for the automatic diag-

nosis in medicine using imaging techniques. These systems, in general, belong to the

diagnostic management of brain pathology. Image technologies also play the key role

in the diagnosis of Pediatric malignancies. CT imaging more than any other imaging

modality provides documentation of the tumor areas, and topography characteristics.

It gives together with additional intravenous infusion of contrast (CT-angiography)

extra data about the blood vessels of the tumor and surrounding tissue.

Now medical experts use vertebrae and ribs to figure out disposition of regions and

organs of human body. The tasks of extraction of spine column, distinguishing and

counting of vertebrae on CT images are usually carried out by medical specialists

manually. This routing work takes significant time and efforts. Algorithms, which

solve or even partly solve mentioned problems, seriously facilitate the work with CT

datasets having from several dozen to hundreds images. A dataset of CT images of a

patient is stored as 2D grayscale scans of body axial section written in the DICOM

format.

A. Zalesky B., M. Nedzved A., V. Ablameiko S. and V. Lukashevich P. (2010).

Automatic Determining of Vertebrae from CT Images.

In Proceedings of the Third International Workshop on Image Mining Theory and Applications, pages 85-91

DOI: 10.5220/0002963000850091

 SciTePress

Medical experts very often need localization of parts of human body and human

organs shown in 2D scans. For this purpose numbering of vertebrae is traditionally

used.

We offer an approach to extract the spine column from CT images and distinguish

vertebrae. It enables automatic segmentation of spine column region and automatic or

semiautomatic separation of vertebrae. We offer our solution in the traditional deci-

sion making form as a prompt for the medical expert who can easily correct the result

of automatic separation of vertebrae.

2 Formulation of Problem

Usually a dataset of CT images of one patient consists of 20÷100 scans. Distance

between scans can vary from 0.1 to 1 centimeter. All images are stored as files in the

DICOM format. Each DICOM file contains 3D coordinates of the top left corner of

the 2D scan image relative to the coordinate system of the CT scanner.

One of difficulties for experts to work with the CT dataset consists in absence of

interconnection of successive numbers of files and their real 3D coordinates. A CT

scan can have arbitrary Z-coordinate relative to the human body Therefore, 2D scans

should be ordered in regard to their space disposition before one begins to segment

human organs.

DICOM pictures are represented as 16-bit grayscale images. Brightness of

DICOM images of 2D CT scans corresponds to the Hounsfield units that characterize

organ densities. Unfortunately, it is not referred to topograms. An appropriate trans-

formation of DICOM images, especially topograms, into the standard 8-bit gray scale

ones can help to get maximum accuracy of further steps. All images are provided

with their 3D coordinates according to the coordinate system of the CT scanner.

Our task is to estimate position of vertebrae and vertebra in a 2D scan. The main

steps of the offered decision rely on finding characteristic vertebra for comparing

different investigation of one patient.

3 Vertebra Segmentation

Vertebra segmentation from CT pictures is an urgent task for many medical applica-

tions. It is widely used to control dynamic of conditions of the spine, to recognize its

deceases and to treat them. Besides, this procedure is an intermediate step for seg-

mentation of abdominal organs, such as the liver, kidneys, and spleen, from CT scan

imagery [1]. Among semiautomatic and automatic approaches to the problem are

model-based, discrete optimization, neural network, active contours, morphologic

methods or their combinations [2-4] with or without of use a prior information etc.

One of the difficulties of vertebra segmentation, often mentioned by authors of algo-

rithms, consists in discrimination between the spine and ribs.

We propose a new automatic algorithm, which allows both: reliable automatic de-

tection of the vertebrae on 2D CT images in DICOM format, and separation of ribs

touching the spine. Also, detection of vertebrae on 256-color gray scale CT images is

possible, as well. The algorithm does not need learning. It contains five main steps:

image preprocessing, finding body section, reliable detection of round part of the

vertebra, accurate segmentation of the bone, which image can break up into several

parts and be touched by ribs, and discrimination between the spine and ribs.

Image preprocessing is performed by version of the region growing algorithm

leaving only the largest connected component of the CT image, which always is the

body section, and removing all other artifacts and noisy clusters. After, the region of

interest (ROI) is found as the bounding box of the body section.

Simple prior information on round shape of a vertebra frontal is used to detect its

location in the image. The shape of matching window is drawn in Fig.1b. Despite of

the simple form of prior information practically all vertebrae were detected by the

chosen window.

Then, top boundary of the vertebrae is determined (Fig.1c). Further outlining of

the vertebra boundary is started from the lowest pixel of the bone (Fig.1d), since 2D

slice CT image of this one piece bone can contain different number of separate piec-

es.

Fig.1. Steps of performance of the algorithm.

The lower boundary of the vertebrae is found after binaryzation of the CT image

at values 1000 -1200 of Hounsfield units corresponding bones of human body. It

allows outlining bottom part of the vertebra, possibly, with touching ribs. Outlined

part of the vertebra without touching rigs is depicted in Fig.1e, and with them – in

Fig.1f.

The algorithm analyzes bottom part of the outlined boundary and recognizes

whether it contains rib frontiers.

In case of recognized touching ribs the algorithm estimates their contact points

with the vertebrae and removes rib frontiers from the bottom boundary. Removed rib

frontiers are shown in Fig. 1g by aquamarine color. The final segmented image with

removes ribs colored in aquamarine can be seen in Fig.1h.

This algorithm was tested by CT-images of children and included in software for

monitoring of mediastinal and retroperitoneal tumors. It is allow to detect description

of organs position by vertebrae geometrical coordinates. This information is very

usefulness for extraction 3D patterns of organs by module-based segmentation algo-

rithms.

4 Definition of Vertebra Orientation

One of the basic feature of analysis CT Images is a topology of organs in body. For

quality topology analysis It is necessary to define body orientation. For solving this

task the algorithm of orientation definition was developed.

Orientation is determined on bone elements and the vertebra by using threshold

segmentation. For the CT image histogram of brightness is divided into three gauss-

like shape classes. On their position we can determine the location of the bodies

(fig.2).

Fig. 2. Histogram of brightness for the CT images.

Position of the last local minimum is determined by the watershed thresholding me-

thod with smoothing of the histogram. Bone fragments carved this thresholding.

However, there is a geometric noise on binary image. These noises correspond to the

soft tissue and acquisition errors. Removal of such errors is performed through sever-

al steps. The first step is to remove small noise, which is performed by analyzing the

size of objects and thin clamped to the edge of objects. This removal is realized by

analyzing the image boundaries (Fig.3).

Fig. 3. Source CT image and Image with remote small object.

Spatial position of bone on the slice is used for definition of bone regions. Practically

all bone elements are placed along the edges. So the overall operation with convex

contour on all elements of the image is used for them to determine. Objects are saved,

if they cross the convex contour.

To determine the remaining elements of the bone dilatation increases convex con-

tour. Analysis of the factor of shape for all elements of the bone tissue determines the

element that corresponding to the vertebra. Then the morphological open operation is

performed with a high depth, which allows to get a round object. Center of gravity is

calculated as a binding element.

Another element of the anchor is the center of gravity for the convex object.

Based on these coordinates calculated line corresponding to the slice orientation

(Fig.4)

Fig. 4. Images from the orientation axis, the arrows indicate the centers of mass of a convex

region and the vertebra.

Further detection of organs regions is carried out based on the watershed. On the

basis of the subtraction of very smoothed watershed lines from the original image and

regions binarization separate organs are defined.

Such analysis allow to define topology properties organs and body that can be

used for monitoring and diagnostic task.

Fig. 5. Regions corresponding to different organs.

5 Conclusions

The presented algorithms allows automatic and semiautomatic extraction of the spine

column and thoracic vertebrae from datasets of CT images stored in the medical

DICOM format. It enables of reliable extraction and separation of thoracic vertebrae

in order to number them relative to the bottom human vertebra with ribs.

In turn, it gives possibility for the medical expert to know the number of the vertebra

he sees in the current 2D scan image.

A new unsupervised algorithm to segment vertebrae in 2D CT images has been pre-

sented. It allows reliable finding this bone and its automatic extraction. The algorithm

does not use prior information on the spine shape. The results of tests showed possi-

bility of automatic extraction of vertebra from CT images. In order to be applicable to

practical extraction the algorithm needs following feasible improvements in order to:

process correctly images, which do not contain the vertebra; process in special way

CT images of patients that took contrast agents; test shapes of outlines contours.

The drawback of the algorithm is its real applicability to separate and number only

thoracic but not lumbar and sacral vertebrae.

Acknowledgements

This work was supported by ISTC project #B-1489.

References

1. Serlie I. W. O., Vos F.M., Truyen R., Post, F.H. van Vliet, L.J., Classifying CT Image Data

Into Material Fractions by a Scale and Rotation Invariant Edge Model, IP(16), No. 12, De-

cember 2007, pp. 2891-2904.

2. Wang L. I., Greenspan M., Ellis, R.E., Validation of bone segmentation and improved 3-D

registration using contour coherency in CT data, MedImg(25), No. 3, March 2006, pp. 324-

334.

3. Seise M., McKenna S.J., Ricketts I.W., Wigderowitz C.A., Learning Active Shape Models

for Bifurcating Contours, MedImg(26), No. 5, May 2007, pp. 666-677.

4. Yao W., Abolmaesumi P., Greenspan M., Ellis R.E., An Estimation/Correction Algorithm

for Detecting Bone Edges in CT Images, MedImg(24), No. 8, August 2005, pp. 997-1010.

5. Jackowski M. and Goshtasby A., "A computer-aided design system for segmentation of vo-

lumetric images," in Handbook of Biomedical Image Analysis, Volume III: Registration

Models, J. S. Suri, D. L. Wilson, and S. Laxminarayan, Kluwer Academic, New York,

2005, pp. 251-272

6. Zagorchev L., Goshtasby A., Satter M., Bernstein T. Segmentation and Registration of Spi-

nal MR and CT Images ( http://www.cs.wright.edu/~agoshtas/wkni_spine.html)