INTRA-PATIENT REGISTRATION METHODS

FOR THORACIC CT EXAMS

José Silvestre Silva

Departamento de Física, Faculdade de Ciências e Tecnologia,Universidade de Coimbra, Portugal

Centro de Instrumentação, Faculdade de Ciências e Tecnologia,Universidade de Coimbra, Portugal

João Cancela

Departamento de Física, Faculdade de Ciências e Tecnologia,Universidade de Coimbra, Portugal

Luísa Teixeira

Clínica Universitária de Imagiologia dos Hospitais da Universidade de Coimbra, Portugal

Keywords: Medical Image Analysis, Image Registration, Lung Tumours, Computed Tomography.

Abstract: Now-a-days CT scanners provide detailed morphological information of pulmonary structures, with great

importance to the diagnostic and follow-up of oncological diseases. When a patient with lung cancer is

submitted to several CT exams during a period of time; these exams need an appropriate registration to

quantify or visualize the tumour’s evolution. We propose a new method for 3D intra-patient registration of

thoracic CT exams and compare its results with several 3D registration methods. The performance of these

registration methods is analysed, computing several normalized figures of merit; we also explore these

metrics to check which is more sensible to changes in CT exams due to the presence of lung tumours. The

results with several cases of intra-patient, intra-modality registration show that the proposed method

provides an accurate registration which is needed for the quantitative tracking of lesions that may effectively

assist the follow-up process of oncological patients.

1 INTRODUCTION

Modern high resolution Computed Tomography

scanners offer more diagnostic options and a better

diagnostic quality. Consequently, it will also

increase the time needed for data reading by the

radiologist. Therefore, computer aid is necessary in

order to increase the level of efficiency and quality

in the diagnostic workflow.

Image registration geometrically aligns two

images: the reference and sensed images. To register

two images it is necessary to find a transformation

so that each pixel in the first image can be mapped

to a pixel in the second (Brown, 1992) (Blaffert &

Wiemker, 2004). The image registration is used in

several clinical scenarios. For instance, consider two

images taken of a patient using different medical

modalities or comparing two CT exams from a

patient, to identify the differences between the two

images in a follow-up study of an oncological

patient. Although this identification can be done by

the radiologist, there is always the possibility that

small, but essential, features could be missed

(Brown, 1992).

In the literature, it is found some work done in

this area. El-Baz (El-Baz, Yuksel, Elshazly, &

Farag, 2005) developed an automatic approach for

the early detection of lung nodules that may lead to

lung cancer. This approach involves performing

rigid registration and then a non-rigid registration to

compensate the lung deformation due to the heart

beats and respiration of the patient; however this

method cannot handle large deformations.

Matsopoulos (Matsopoulos, Mouravliansky,

Asvestas, Delibasis, & Kouloulias, 2002) proposed

an automatic elastic registration scheme applied on

thoracic CT exams of patients diagnosed with non-

small cell lung cancer.

285

Silvestre Silva J., Cancela J. and Teixeira L. (2009).

INTRA-PATIENT REGISTRATION METHODS FOR THORACIC CT EXAMS .

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing, pages 285-290

DOI: 10.5220/0001538002850290

 SciTePress

Volumetric warping and registration of CT lung

volumes have been investigated by Li (Li,

Christensen, Dill, Hoffman, & Reinhardt, 2002),

whose approach uses point correspondence of

landmarks that are expanded over the entire volume

by means of an iterative method. Although this

method has shown good results for mapping lung

deformation due to respiration, it requires the

manual registration of landmarks.

Betke (Betke, Hong, Thomas, Prince, & Ko,

2003) developed an automated method for

registering CT images of the chest; it detected

anatomical landmarks: the trachea, sternum and

spine, then he used an iterative lung surface

registration based on minimizing Euclidean

distances. The locations of the pulmonary vessel

branch points and nodules were manually defined.

Blaffert (Blaffert & Wiemker, 2004) studied the

precision and computation time of a rigid body using

an affine and a spline based elastic registration

approach on the full data volume; he compared the

results to an affine registration that was preceded by

a segmentation of the lung.

Boldea (Boldea, Sarrut, & Carrie, 2005)

investigated the deformable registration methods for

a breath-hold reproducibility study in radiotherapy,

analysing internal lung residual motion between

several 3D CT scans taken from the same patient, at

the same level of the breathing cycle.

West (West, Maurer, & Dooley, 2005) examined

the problem of deformable registration of the

abdomen and was interested in modelling respiratory

motion of abdominal organs, because the

deformation of the lungs during the respiratory cycle

can lead to the movement of others organs (liver,

kidney, etc.); he used twenty-one landmarks selected

manually.

Chambon (Chambon et al., 2007) presented a

CT-PET landmark-based registration method that

uses a breathing model to guarantee physiologically

plausible deformations.

Fung (Fung, Wong, Cheng, Grimm, & Uematsu,

2005) compared two image fusion techniques for the

localization of patient position during radiation

release for cancer patients.

Tang (Tang, Hamarneh, & Celler, 2006)

presented an automatic and accurate technique for

3D registration of SPECT and CT, which allowed

the attenuation correction of SPECT images and the

fusion of the anatomic details from CT and the

functional information from SPECT.

Ruan (Ruan, Fessler, Robserson, Balter, &

Kessler, 2007) studied a method that takes into

account different types of tissues, especially bone, in

non-rigid registration. Chen (Chen, Varley, Shark,

Shentall, & Kirby, 2007) presented a 3D-2D image

registration algorithm for pre-treatment validation in

radiotherapy.

In the present work, we propose a methodology

for the 3D intra-patient registration of thoracic CT

exams. We compare performances analysing the

processing time and the values of similarity metrics

for each method. We also studied the behaviour of

several normalized similarity metrics in the presence

of pulmonary tumours in oncological patients.

2 METHOD

In this paper, we present a method for the

registration of pulmonary CT exams and compare its

performance with traditional registration method and

also with two optimised registration methods. Also

in this work, we search for the best metric, sensible

to changes in CT exams, due to the presence of lung

tumours.

2.1 Pre-processing

In the high resolution CT exams, the images are

sensitive to noise, especially in the extra-thoracic

region, where there is air. As noise can contribute

negatively to the lung’s segmentation, before any

processing, and after comparing several denoising

filters (J. S. Silva, Silva, & Santos, 2003), the noise

is attenuated using a geometric mean filter (Sonka,

Hlavac, & Boyle, 1998).

The pulmonary regions are identified using a

previous developed (A. Silva, Silva, Santos, &

Ferreira, 2001) (J. S. S. Silva, 2005) and validated

algorithm (Santos, Ferreira, Silva, Silva, & Teixeira,

2004) producing always one contour for each lung.

This algorithm uses information from the CT image

histogram and, with a chain of morphological

operations, identifies the left and right lung contours.

Even when the lungs are visually merged, two

contours are always identified, defining the frontier

of left and right lungs. After pulmonary

segmentation, all binary slices are joined to create

one 3D pulmonary image.

Finally, in this pre-alignment step, it is computed

the mass center of both 3D CT exams and performed

a translation on the second exam, in order to both

mass centers become coincident.

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

286

2.2 Affine Registration

The registration can be understood as the

determination of spatial alignment between images.

For the present method, we consider affine

transformations, as the only relevant transformation

(Blaffert & Wiemker, 2004).

2.2.1 Transformations

We defined the following transformation matrix: T

for translation on x, y ,z; R

, R

for rotation over

3 axis, C

, C

and C

for shearing on planes xy, xz

and yz and S for scaling. The various transformation

matrices are multiplied among them to obtain a

global matrix, in order to process all voxels from the

CT exam:

SCyzCxzCxyRzRyRxTGLOBAL ×××

×××=

(1)

In traditional 3D registration methods, all

translations are iteratively searched, and for each

translation all rotations are searched, and so on;

processing a 3D image becomes a very long process.

In order to overcome this limitation, we propose a

registration method that sequentially performs each

transformation.

In this proposed method, the first step is the

search for the best translation, then holding this

value; it searches for the best rotation over x axis.

Holding these two best values; it searches for the

best rotation over y axis, and so on, obtaining the

best values for each transformation in a much faster

approach than in traditional 3D registration methods.

In the second step, it uses the best values found in

the previous step as starting point, and then repeats

the same procedure described in the first step,

searching for new best values.

2.2.2 Boundaries

Setting boundaries for transformation is very

important to reduce the processing time. For a CT

exam that has been correctly acquired, it should not

have a displacement of more than ¼ of width of the

exam, and the patient body should not have a

rotating higher that 15º on the table, otherwise the

examination is considered inappropriate, because the

anatomical structures may exceed the limits of the

image; these are the limits used for the

transformations.

To enhance the speed of our method, we start by

computing the width of the CT exam and use 1/8 of

the width as the step for searching the best

translation, over the x, y and z axis in a range

from −¼ of the exam width, to +¼ of the exam

width. After searching for these axes, the best value

is found comparing all values computed for the

similarity metric.

The best translation is identified and using this

value as the starting point, a new search is

performed in a range from −1/8 of the exam width,

to +1/8 of the exam width centred at the starting

point, using a step of 1/16 of the exam width, which

is half of the previous used step. This procedure is

repeated until the step reaches the unit.

Then we search for the best rotation, following a

similar procedure used for translation. Holding the

translation best value found, we search for the best

rotation, along the 3 rotation axis, in a range of

[−15º; +15º] with a step of 7.5º. After identifying the

best rotation, we use this value as the starting point

and search for a new best value in a range of ± 7.5º,

centred at the starting point, with a step of 7.5º/2.

This procedure is repeated until the rotation step

reaches the value of one degree. Similar procedures

are used for scale and for shear.

2.3 Metrics

We used several normalized similarity metrics, also

known as figures of merit to quantify the differences

between the reference exam and the exam under

analysis. These three metrics: Sum of the Absolute

Differences (SAD), Correlation (R) and Normalized

Mutual Information (NMI) have values between 0

(for two different images) and 1 (for two coincident

images) (Fitzpatrick, Hill, & Calvin R. Maurer,

2000) (Hill, Batchelor, Holden, & Hawkes, 2001)

(Pratt, 2001). The Mutual Information (MI) metric is

shown to help computing the NMI.

∑∑∑

−−−−−=

oknjmiBkjiA

BASAD ),,(),,(

1),(

(2)

∑∑ ∑∑∑∑

∑∑∑

−−−×

oknjmiBkjiA

),,(),,(

(3)

∑∑

∈∈

AaBb

baBAMI

)()(

),(

log),(),(

ρρ

(4)

)()(

),(2

),(

BHAH

BAMI

BANMI

with

[]

∑

∈

−=

aaAH )(log)()(

ρρ

(5)

where, A and B are images; N is a normalization

coefficient; i, j and k are the coordinates on the

image; m, n and o are the displacement values to the

reference image; ρ

(a,b) is the joint probability of

INTRA-PATIENT REGISTRATION METHODS FOR THORACIC CT EXAMS

287

the image and ρ

(a), ρ

(b) are the probability of

images A and B, respectively.

2.4 Other Registration Methods

Two registration methods, using the Simplex

algorithm and the Patter Search algorithm, were also

implemented to compare their performance with the

proposed method.

The Nelder-Mead Simplex algorithm is a direct

search method for multidimensional unconstrained

minimization. Without any derivative information, a

scalar-valued nonlinear function of n real variables

using only function values is minimized. The

Nelder-Mead algorithm preserves at each stage a

nondegenerate simplex, a volume different from

zero in n dimensions which is the convex surface of

n+1 vertices. This method starts with a simplex,

specified by its n + 1 vertices and the related

function values for each iteration. At least one test

point is calculated, as well as their function values,

and the iteration ends with the levels sets delimited

(Lagarias, Reeds, Wright, & Wright, 1998).

The Pattern Search algorithm is also a direct

search method and uses the function from a

prearranged pattern of points fixed around the

current best point, using shifts that guarantee

determined minimal conditions in order to ensure the

strong performance of the method. This process is

repeated with the pattern centred on the new best

point whenever certain minimal conditions are

ensured. In other words, the reduction of the size of

the pattern occurs and the function is sampled once

again. The goal of the Pattern Search is sampled at

set points which are broader than in the Simplex-

based methods (Torczon, 1997) (Lewis & Torczon,

2002).

3 RESULTS AND DISCUSSION

In this section, we present the results from the

comparison of our method and three other methods:

the traditional registration method and two other

methods with optimization algorithms: the Simplex

algorithm and the Pattern Search algorithm.

In a second step we used exams from an arbitrary

patient and perform the registration of all exams, to

analyse the behaviour of normalized similarity

metrics.

The results of all registration methods were

computed on a desktop computer Intel Core 2 Quad,

4GB RAM, using Matlab.

Our dataset has 40 CT exams from 10 patients,

each exam has about 100 sections, with 512×512

pixels, a resolution of 0.781×0.781×5mm

and each

section is adjacent to its neighbours.

3.1 Comparing Registration Methods

We compare the results of several 3D registration

methods, performing the intra-patient registration of

two exams, acquired with one month interval.

In table 1 it is shown the results of 3D intra-

patient registration of pulmonary CT exams,

downsampled to 128×128×n (where n is the original

number of sections), which include correlation

values (initial value, after pre-processing /

preliminary alignment based on mass center, final

value), processing time and the number of iterations.

Table 1: Correlation values: registration of two exams

from patient A.

(Patient A) Correlation values Time No. of

Method: Initial Pre-align Final (min) iterations

Traditional 0.854 0.857 0.939 1830 21870

Our method 0.854 0.857 0.904 20 234

Pattern Search 0.854 0.857 0.929 116 1385

Simplex 0.854 0.857 0.878 12 128

From table 1, we see that the processing time of

our method and Simplex method are much lower

than the other two methods, which suggests that our

method is a fast 3D registration technique, even

when compared with a method that uses an

optimization algorithm (Pattern Search method). We

also observe that the best correlation values are

obtained with the Pattern Search method (and with

traditional method) and the worst value, with the

Simplex Algorithm. So, the traditional method, due

to the long processing time, is rejected.

In table 2 it is shown the results of 3D intra-

patient registration of pulmonary CT exams,

downsampled to 64×64×n (as described for table 1)

and the initial / pre-processing / final values, for

Normalized Mutual Information.

Table 2: Normalized Mutual Information values:

registration of two exams from patient B.

(Patient B) Normalized Mutual Infor. Time No. of

Method: Initial Pre-align Final (min) iterations

Traditional 0.114 0.628 0.799 1050 21870

Our method 0.114 0.628 0.705 12 234

Pattern Search 0.114 0.628 0.796 45 1579

Simplex 0.114 0.628 0.681 15 286

In table 2, we see that our method and the

Simplex method accomplish the lower processing

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

288

time. However, the normalized mutual information

achieved by Simplex method is worse than the value

obtained by our method, which suggests that our

method is one of the best, in comparison with the

three methods.

3.2 Metrics in Tumour Discrimination

Using six exams (A1, A2, A3, A4 A5 and A6)

acquired during six months, one exam in each

month, for the patient A (patient with lung cancer

undergoing intensive therapy), we performed the

registration using our method, of first exam A1 with

second exam A2, also the first exam A1 with third

exam A3, and so on, as shown in table 3. For each

registration, the SAD, R and NMI metrics were

computed. To reduce computational time, the exams

were downsampled to 128×128×n and the lungs

were segmented, producing 3D binary volumes,

corresponding to the pulmonary regions.

Table 3: Intra-patient registration method, using 6 CT

exams from a patient with lung cancer.

SAD

A1 w/ A2 A1 w/ A3 A1 w/ A4 A1 w/ A5 A1 w/ A6

Initial 0.937 0.891 0.924 0.817 0.900

Pre-align 0.939 0.923 0.949 0.936 0.929

Final 0.961 0.956 0.969 0.961 0.956

A1 w/ A2 A1 w/ A3 A1 w/ A4 A1 w/ A5 A1 w/ A6

Initial 0.854 0.742 0.831 0.598 0.766

Pre-align 0.857 0.817 0.886 0.860 0.834

Final 0.903 0.891 0.930 0.914 0.896

NMI

A1 w/ A2 A1 w/ A3 A1 w/ A4 A1 w/ A5 A1 w/ A6

Initial 0.588 0.387 0.525 0.191 0.417

Pre-align 0.594 0.512 0.640 0.584 0.537

Final 0.692 0.665 0.752 0.708 0.670

Using the data from table 3, we create a group

for each metric and produce a box-plot graphic, to

observe the dispersion of metric values.

Figure 1: Dispersion values of similarity metrics.

From figure 1, and discarding the three outlier

points, we see that SAD has the lower dispersion

interval (values from 0.891 to 0.969) and the NMI

has the higher dispersion interval (values from 0.387

to 0.752) which suggests that NMI is more sensible

to the presence of tumours.

These results were confirmed with results from

intra-patient registration of CT exams, using other

patients with lung cancer.

4 CONCLUSIONS

In this paper, we addressed the problem of

registering volumetric pulmonary CT exams of

patients with lung cancer. We propose an automatic

3D intra-patient registration method. It starts by

segmenting the lungs and building a 3D binary

image of the pulmonary region. The centre of mass

is computed and the exams are coarsely aligned.

Then, a 3D registration is performed using a

downsampled volume from the original 3D image.

The performance of our method is compared with

the traditional registration method and also with two

optimised methods and we conclude that our method

is the best compromise between processing time and

similarity metric values.

Also, we compare the results of several

normalized similarity metrics used in the 3D

registration of CT exams and conclude that

normalized mutual information is the metric more

sensible to the changes in CT exams due to the

presence of lung tumours.

The results with several cases of intra-patient,

intra-modality registration show that this method

provides accurate registration which is needed for

the quantitative tracking of lesions that may

effectively assist the follow-up process of

oncological patients.

REFERENCES

Betke, M., Hong, H., Thomas, D., Prince, C., & Ko, J. P.

(2003). Landmark Detection in the Chest and

Registration of Lung Surfaces with an Application to

Nodule Registration. Medical Image Analysis, 7(3),

265-281.

Blaffert, T., & Wiemker, R. (2004). Comparison of

different follow-up lung registration methods with and

without segmentation. SPIE Medical Imaging 2004,

5370, 1701-1708.

Boldea, V., Sarrut, D., & Carrie, C. (2005). Comparison of

3D Dense Deformable Registration Methods for

Breath-hold Reproducibility Study in Radiotherapy.

INTRA-PATIENT REGISTRATION METHODS FOR THORACIC CT EXAMS

289

SPIE Medical Imaging: Visualization, Image-Guided

Procedures, and Display, 5747, 222-230.

Brown, L. G. (1992). A Survey of Image Registration

Techniques. ACM Computing Surveys, 24(4), 325-376.

Chambon, S., Moreno, A., Santhanam, A. P., Rolland, J.

P., Angelini, E., & Bloch, I. (2007). CT-PET

Landmark-based Lung Registration Using a Dynamic

Breathing Model. Proceedings of the 14th

International Conference on Image Analysis and

Processing (ICIAP 2007) - Volume 00, 691-696.

Chen, X., Varley, M. R., Shark, L.-K., Shentall, G. S., &

Kirby, M. C. (2007). Automatic 3D-2D image

registration using partial digitally reconstructed

radiographs along projected anatomic contours. Proc.

of International Conference on Medical Information

Visualisation - BioMedical Visualisation (MediViz

2007), 3-8.

El-Baz, A., Yuksel, S., Elshazly, S., & Farag, A. (2005).

Non-rigid registration techniques for automatic

follow-up of lung nodules. Proc. of Computer Assisted

Radiology and Surgery (CARS'05), Berlin, Germany,

1115-1120.

Fitzpatrick, J. M., Hill, D. L. G., & Calvin R. Maurer, J.

(2000). Image Registration. In M. Sonka & J. M.

Fitzpatrick (Eds.), Handbook of Medical Imaging

(Vol. 2, pp. 447 - 514): SPIE Press.

Fung, A. Y. C., Wong, J. R., Cheng, C.-W., Grimm, S. L.,

& Uematsu, M. (2005). A comparison of two image

fusion techniques in CT-on-Rails localization of

radiation delivery. Physica Medica, 21(3), 113-119.

Hill, D. L. G., Batchelor, P. G., Holden, M., & Hawkes, D.

J. (2001). Medical image registration. Physics in

Medicine and Biology, 46(3), 1- 45.

Lagarias, J. C., Reeds, J. A., Wright, M. H., & Wright, P.

E. (1998). Convergence Properties of the Nelder-Mead

Simplex Method in Low Dimensions. SIAM J. on

Optimization, 9(1), 112-147.

Lewis, R. M., & Torczon, V. (2002). A Globally

Convergent Augmented Lagrangian Pattern Search

Algorithm for Optimization with General Constraints

and Simple Bounds. SIAM J. on Optimization, 12(4),

1075-1089.

Li, B., Christensen, G. E., Dill, J., Hoffman, E. A., &

Reinhardt, J. M. (2002). 3D inter-subject warping and

registration of pulmonary CT images for a human lung

model. SPIE - Medical Imaging 2002: Physiology and

Function from Multidimensional Images, 4683, 324-

335.

Matsopoulos, G. K., Mouravliansky, N. A., Asvestas, P.

A., Delibasis, K. K., & Kouloulias, V. (2002).

Thoracic non-rigid registration combining self-

organizing maps and radial basis functions. Med Phys.,

29(2), 201-213.

Pratt, W. K. (2001). Digital Image Processing: John Wiley

& Sons, Inc.

Ruan, D., Fessler, J. A., Robserson, M., Balter, J., &

Kessler, M. (2007). Nonrigid Registration with

Regularization Incorporating Local Tissue Rigidity.

Phys. Med. Biol. (in revision).

Santos, B. S., Ferreira, C., Silva, J. S., Silva, A., &

Teixeira, L. (2004). Quantitative Evaluation of a

Pulmonary Contour Segmentation Algorithm in X-ray

Computed Tomography Images.

Academic Radiology,

11(8), 868-878.

Silva, A., Silva, J. S., Santos, B. S., & Ferreira, C. (2001).

Fast Pulmonary Contour Extraction in X-ray CT

Images: A Methodology and Quality Assessment.

SPIE - Medical Imaging 2001: Physiology and

Function from Multidimensional Images, 4321, 216-

224.

Silva, J. S., Silva, A., & Santos, B. S. (2003). Denoising

Methods in High-Resolution X-Ray Computed

Tomography. Proceedings of BioEng 2003 - 7th

Portuguese Conference on Biomedical Engineering,

59 (extended version in CD of Proceedings).

Silva, J. S. S. (2005). Segmentação Pulmonar em Estudos

de Tomografia Axial Computorizada: PhD Thesis (in

Portuguese), Universidade de Aveiro.

Sonka, M., Hlavac, V., & Boyle, R. (1998). Image

Processing, Analysis, and Machine Vision (2nd ed.):

PWS Publishing.

Tang, L., Hamarneh, G., & Celler, A. (2006). Co-

registration of Bone CT and SPECT Images Using

Mutual Information. IEEE Symposium on Signal

Processing and Information Technology, 116-121.

Torczon, V. (1997). On the Convergence of Pattern Search

Algorithms. SIAM J. on Optimization, 7(1), 1-25.

West, J. B., Maurer, C. R., & Dooley, J. R. (2005). Hybrid

point-and-intensity-based deformable registration for

abdominal CT images. Proceedings of SPIE Medical

Imaging 2005, 5747, 204-211.

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

290