Organ segmentation and absorbed dose calculation

The software of GAMOS version 5.1.0 can read DICOM RTSTRUCT file to calculate the dose or energy deposited in a structure. Before being used, they have to be converted into the list of CT voxels that belong to each structure. One of the tools that can be used to draw structures on a CT image file is Carimas.

ORGAN SEGMENTATION AND ABSORBED DOSE

CALCULATION

Nguyen Thi Phuong Thao 1 , Nguyen Thien Trung 2

1 Vietnam Atomic Energy Institute, 2 Techbase-Yahoo Japan The software of GAMOS version 5.1.0 can read DICOM RTSTRUCT file to calculate the dose or energy deposited in a structure Before being used, they have to be converted into the list of CT voxels that belong to each structure One of the tools that can be used to draw structures on a CT image file is Carimas We have developed a tool that could identify the CT voxels that belong to the structures drawn by Carimas and could write this information in a file usable by GAMOS As a first check of the tool, A geometrical comparison of the areas calculated with Carimas with those calculated by GAMOS using the converted file has been made In a second step, using a source of F - 18 placed in the water volumes of the NEMA phantom and comparing the absorbed dose on a disintegration in the drawn structures calculated by Carimas with those calculated by GAMOS have been implemented After validating the tool, using the CT images of 5 adult male patients, drawing in Carimas the structures defining several structures, i.e kidney and spleen, and calculating the self-absorbed and cross-absorbed dose with GAMOS were made A patient to patient variability of up to 41.75% (spleen → kidney) was found

I INTRODUCTION

Keywords: DICOM RTSTRUCT, GAMOS, Geant4, Carimas, voxel phantom.

The information on the anatomy, the

activity, and the dose distributions in different

parts of human body (whether corresponding

to real organs or to simulation issues) is crucial

for medical staff to determine the appropriate

nuclear medicine or radiotherapy treatment,

as well as to understand the treatment effects

One of the tasks that need this information is

the assessment of the possible side effects of

internal exposure to radionuclide on the organs

and tissues that strongly affect to the health of

patients [1] These features are imbedded and

enable to be extracted from the CT image

Currently, plenty of different commercial

software packages can now be used to draw the structures on a CT image [2] and allow for the reading of those structures and calculation

of the activities and deposited doses of treatments [3] The volume of interest can be drawn by many contours on each z slice, or even in three-dimensional consideration For instance, 3D Slicer is an open source software platform for medical image informatics, image processing, and three-dimensional visualization [4] It is appreciated for its easy use Similarly, ITK - SNAP is a software used to segment structures in 3D medical images with many utilities [5] They can link cursors for seamless 3D navigation with manual segmentation in three orthogonal planes at once The software supports many different 3D image formats, including NIfTI and DICOM Compared to other larger, open-source image analysis tools,

Corresponding author: Nguyen Thi Phuong Thao,

Vietnam Atomic Energy Institute

Email: nguyen.thi.phuong.thao.8488@gmail.com

Received: 27/11/2018

Accepted: 12/03/2019

ITK-SNAP design focuses specifically on the

problem of image segmentation It was written

only for 64 bit modes Developed for both CT

and PET images, Carimas is a general medical

imaging processing platform developed in

Turku PET Centre, Finland [2] Carimas can

read only uncompressed images and run only

on Windows After loading the PET/CT images

to Carimas, it allows users to define a structure

by drawing a volume or a set of contours on

each slice Information about coordinates of

points forming the structure can be saved as a

VRML file Most of the software can measure

area, mean, standard deviation, min and max

of selection or entire image, measure lengths

and angles, generate histograms and profile

plots, calculate the activity or HU values on the

drawing

However, the general drawback of these

software is that they do not support the feature

of outputting data to indicate which voxels

belong to a volume Besides, there is no free

available software that can be used to draw

the structures on a CT image, identify to which

structures each of the CT image voxel belongs,

and calculate the absorbed dose in them by

using the Monte Carlo simulation method

proved to be the most precise way to do such

calculations

GAMOS is an easy-to-use Monte Carlo

simulation tool built on the top of the Geant4

toolkit [6] Besides image processing and

3D visualization, GAMOS can make use of

DICOM CT structures to compute the absorbed

dose or energy deposit in each structure, a

very important task in nuclear medicine To

do this, GAMOS offers a simple text format

to list the structures that each image voxel

belongs to, plus a robust mechanism to identify

this structure on the voxel map after reading the structural information from a DICOM RTSTRUCT file

In this report, we develop a tool that uses the drawing on CT images (VRML file) of Carimas for specifying the structure to which the image voxels belong The information of these voxels can be converted to structure text format of GAMOS in order to calculate the absorbed dose on a disintegration (the S value) at voxel level

II METHODS

1 Study setting

In this section, a procedure to segment a structure and extract the information from CT images is proposed Such information is vital for calculation of absorbed dose at voxel level using GAMOS

2 Materials and methods

Firstly, the desired structures have to be obviously indicated and embedded in the CT images This step is straightforwardly resolved using the Carimas program [2] Once Carimas inputs the DICOM CT images, the images are displayed as both 2D and 3D images The users can use the default shapes of Carimas such as sphere, cylinder, tube, and cube Subsequently, the editing functions of Carimas enable one to modify the Volume of Interest (VOI) for getting the desired shape as illustrated

in Figure 1 The VOI is then saved as a VRML file, containing the coordinates of the points on the surface of the VOI The information on the coordinates of these points are used to figure out which voxels belong to the VOI in the next stage using the so-called flood-fill algorithm as described in the following [7 - 8]

Figure 1 Arbitrary structures drawn by Carimas on top of CT images The closed, blurred

pink shape indicates the desired VOI.

The flood-fill algorithm is widely used

but not for such purposes In this article, we

adopted this algorithm to precisely indicate

which voxels belong to a structure taken from

the previous phase of procedure A box called

the World Volume (WV) is firstly created The

WV adequately contains the VOI as presented

in Figure 2 in which the black circles covered

by the pink square is our desired structures

All appropriate voxels of the CT images in WV

are then figured out including their coordinate

information Note that such a performance is

straightforward due to the available minimum

and maximum values of x, y, z coordinates

saved in the VRML files The coordinates of all

the voxel vertices in the WV are used to specify

all the voxels containing the points locating on

the surface of the VOI It is also noted that

those voxels cannot cover the surface of the

VOI In the VRML file, the points on the surface

of the VOI are listed line-by-line Each line

corresponds to the points located on the same

plane, three of them are arbitrarily chosen to create a pediment On each pediment, each edge is divided into several segments so that the dimension of each part is smaller than that

of the voxel Then each vertex is connected

to its neighbors on the same pediment to look for all voxels on that line The loop is repeated for all pediments and planes to get a voxel net covering the surface of the VOI

After covering the VOI, a “prototype” voxel

is embraced and set free to flood in the VOI until it touches the VOI’s surface All voxels that are in contact with that prototype one are then chosen Similarly, all of the chosen voxels are again moved until they fill the whole VOI The information of those specified voxels in the VOI

is saved to the second VRML file which can later be recalled by Carimas The geometry

of the treated volume is then compared to the original VOI for validating results The smaller the voxel size is, the more consistent the treated volume is to the shape of the VOI

Figure 2 Illustration of specification of the WV strictly covering the VOI The upper left, up-per right, and lower left panels are the WV in (Oy, Oz), (Ox, Oz), and (Ox, Oy) plains,

respec-tively The lower right panel shows three dimensional WV.

Once the information of structures is fully

available, it has to be transformed to GAMOS

format The structures from a DICOM file

of RTSTRUCT modality can be converted

to the simple text format of GAMOS For

that purpose, one adds the name of the

DICOM in the metadata file containing the

information of the DICOM CT images used by

the DICOM2G4 command After reading the

structure’s information from the RTSTRUCT

file, GAMOS identifies on each (Ox, Oy) plain

of the CT image which voxels belong to each

structure provided that they are surrounded by

the structure

From this step, GAMOS is consistently

used for calculating the S values As presented

in MIRD 11 [9], to estimate the absorbed dose

to a structure, the following two components

are necessary: (1) the time-integrated activities

(Bqs) which correspond to the number of

disintegrations occurring in all source structures and (2) the S values of absorbed dose per unit cumulated activity (mGy(Bqs) - 1) The S values depend on the absorbed fraction, which are computed by the phantom with a Monte Carlo transport method:

Where mtarget is the target tissue mass, ∆i

is the equilibrium dose constant for particles of

a particular type and energy, here indicated by

i, and ϕ_i represents the absorbed fraction of energy for target structure for particle i emitted

in the source structure From the CT images, GAMOS can calculate the dose or energy deposit in each voxel With DICOM RT structure

in the g4dcm file, GAMOS can calculate the dose/energy/S value of each structure The S values are calculated for the phantom and for

S(target←source) = ∑∆i ∅i(target←source)

m target

the specific patient In this article the S values

for the patients are calculated for kidney and

spleen

The S values are calculated for both NEMA

phantoms [10] and 5 adult male patients The

three water spheres with the radii of 6.20mm,

7.82mm, 13.36mm are filled with F-18 These

spheres do not overlap With each sphere, the

CT images are cropped to get a cube which

is appropriate to contain the sphere The voxel

dimensions are equal to 1/20, 1/40 and 1/60 of

the diameter of the spheres, respectively The

S values and the difference in volume of each

sphere are compared with those of OLINDA/

EXM [11] In real situation, the CT images of

5 adult male patients (on Siemens scanners)

are converted to GAMOS text files (g4dcm) to

get the S values for kidneys and spleen The

voxel dimension on three axes are 0.97mm,

0.97mm, and 2mm Note that the sources are

assumed to be uniform distributed by F-18

In all calculation, the number of events is

100 million, the data of the isotopes is taken

from the LUND database The S values of

the patients will then be used to calculate the

biggest variability from patient to patient

3 Research ethics

Following the research ethics, the study is

based on PET / CT images of patients, provided

by Cho Ray Hospital Research has no effect

on patients’ health and patient information is

kept strictly confidential

III RESULTS

To evaluate the applicability and validity

of established algorithm, all crucial quantities

indicated by this algorithm including the shape, position, and volume of the voxel set are firstly compared with those directly obtained

by Carimas CT images of an adult male are used in which the VOIs represent a spherical tumor, left kidney, right kidney, and spleen The selected sizes of each voxel are: 5mm, 3mm, 2mm and 1.5mm Figure 3 shows the depiction of the mentioned structures by sets

of appropriate voxels determined by flood fill algorithm

Figure 3 indicates that the shape and the location of the structures are relatively well reflected by the sets of voxels, especially within the area constrained by solid curve showing the boundary of the structures However, at the border, the voxelization of these structures creates an inevitable difference in shape This difference is quantitatively interpreted by their volumes and stems from the size of individual voxel and can efficiently be reduced as the size

of voxels decreases as shown in Table 1 Here the volumes of voxel sets are compared with those of structures given by Carimas as the size of voxel is curtailed from 5mm to 1.5mm

As decreasing the size of voxels, the volume differences are also remarkably reduced by around a factor of ten For smallest voxel size

of 1.5mm, such differences are less than 5% for all examined cases Note that the case of tumor sphere exhibits the highest deviation This feature is easily understandable since

it is the smaller structure and has a perfectly rounded shape, which challenges the interpretation when using square voxels

Figure 3 Several representative human structures indicated by sets of appropriate voxels.

Table 1 The comparison of volumes of the structures and the voxel set

for difference size of voxel

Structure Voxel dimension

(mm)

Volume of structure (mm 3 )

Volume of the voxel set (mm 3 )

Difference (%)

|Vset of voxel-Vstructure|

Vstructure

Right kidney

5 3 2 1.5

211006

252250 232767 221296 215861

19.54 10.31 4.87 2.30

Left kidney

5 3 2 1.5

226776

274500 248373 237032 231096

21.04 9.52 4.52 1.90

Spleen

5 3 2 1.5

214906

286000 247050 230168 222594

33.08 14.96 7.10 3.58

Sphere tumor

5 3 2 1.5

33491

48125 39042 36304 34742

43.69 16.57 8.40 3.74

The voxel size can alternatively be reduced to the resolution of the CT image (less than 0.5mm), which may well reflect the shape and volume of the volume of interest To determine which voxel belongs to a structure, it is based on the voxel center position If the center of a voxel is in a structure, the whole voxel belonging to this structure will be considered This method has a crucial drawback in application to very thin organs such as skin and mucosa due to large deviations as shown in Table 1 for a spherical tumor

The consideration of S values will be carried out below, firstly for NEMA phantom

- S values of NEMA phantom

The dose for water spheres calculated by GAMOS and OLINDA/EXM is presented in Table 2

F - 18 is assumed to be uniformly distributed in the water sphere, emitting isotropic radiation The comparison of the self-dose of the sphere will be done, when directly calculated with the spherical model using the OLINDA/EXM software and when calculated from the dose of each voxel calculated

by GAMOS In all case, the greater the volume difference, the greater the difference in S value The difference in S value is less than 5% when the difference in volume is less than 10% and when the voxel size is equal to 1/20 of the phantom's diameter

Table 2 The S values (mGy/MBq) for water spheres of NEMA phantom filled with F - 18.

The difference of this dose is due to the difference at the surface of the sphere A voxel at the surface of the sphere has a part of volume inside the sphere Taking the energy and the volume of whole surface-voxel will cause differences when calculating the average dose for the sphere

- The S value for the organs of patients

Using the established tool and GAMOS software to calculate the S values for the four important organs of five adult patients is carried out (Table 3 and Table 4) We assume that the

Mass

(g)

Voxel

Size

(mm)

Number

of voxel

in sphere

Volume

of the sphere (mm 3 )

Volume of the group

of voxels (mm 3 )

S value calculated

by GAMOS

S value calculated

by OLINDA

Difference of volume (%)

V GAMOS -V OLINDA EXM⁄

V OLINDA EXM⁄

Difference of S value (%)

S GAMOS -S OLINDA EXM⁄

S OLINDA EXM⁄

m = 1

0.62 5136

1000

1224.05 3.41E - 2

3.944E - 2

m = 2

0.78 5160

2000

2448.68 1.75E - 2

2.019E - 2

m = 10

1.34 4966

10000

11948.71 3.66E - 3

4.27E - 3

radiopharmaceutical is uniformly distributed and isotropic in the source organ The S value is calculated by the Monte Carlo method with the number of event so that the statistical error is kept below 5% We see the results of the different patients are varied The biggest differences (patient

to patient) in radiation exposure in these cases are up to 41.75% for cross-dose and 22.14 % for self-dose

Table 3 S values for the kidneys and spleen in the adult male patients for F - 18 assumed

to be uniformly distributed in kidney (mGy/MBq)

Target structure

difference (%)

Table 4 S values for the kidneys and spleen in the adult male patients for F - 18 assumed

to be uniformly distributed in the spleen (mGy/MBq)

Target structure

difference (%)

Smax-Smin

Smin

Smax-Smin

Smin

IV DISCUSSION

In the Figure 3 and Table 1, the shape

and the congregation of the voxels present

the shape and the position of the structures

well The smaller the voxel size, the closer the

shape and the volume of the voxel set is to the

shape and volume of the structure

From the S value of the water spheres of

the NEMA phantom, seeing that in the selected

voxel dimensions, the S value calculated

by GAMOS is always smaller than the result

calculated by OLINDA/EXM The smaller the

size of the voxel, the smaller the difference

between the results obtained by GAMOS and

OLINDA/EXM With a voxel size of 1/60 the diameter of the sphere, the difference of S value is less than 5% when the difference of volume is less than 10% With the S values

of the adult male patients, seeing that with all pairs of source-target organs, the S values of different patients are different In these cases that have the biggest difference, the source and the target are not the same A patient to patient, variability of up to 41.75% has been found when the S values are spleen → kidney This is caused by differences in the anatomy and the distance between the structures

Many geometric phantoms as well as voxel

phantoms are used to represent the anatomy of

human subjects The S values for the phantom

will be used along with the residence time of

the pharmaceutical in the body to calculate

the accumulate dose for the patients Besides,

some authors assumed that the S value from

structure A to structure B is equal to the one

of the structure B to A (S(A←B) = S(B←A))

[11] Based on the above results, it is found

out that, the above expect cannot be achieved

Therefore, this assumption for calculating the

S values for these pairs of structures cannot

be used

In fact, the distribution of radioactivity and

density in a body is very complex; it may

cause an even bigger difference in the dose

distribution Therefore, it is very important to

calculate the dose for each patient This has

not been done in Vietnam and many countries

in the world To estimate the absorbed dose

for a patient, it can use PET/CT images It can

help to calculate the absorbed dose for each

voxel And the last step is determining which

voxel belongs to a structure, in order to best

estimate the average dose that each structure

receives

V CONCLUSION

Because the distribution of radioactive

pharmaceuticals in the body and the anatomy

of patient is very complex, it is important to

calculate the absorbed dose for each patient

GAMOS is the free and easy to use It can

use PET/CT images to calculate the absorbed

dose at voxel level for patients First, it needs

to convert both DICOM PET and CT images

to the GAMOS format The PET image gives

the source distribution and CT images give

anatomical information of each structure Then

the tool helps find out which voxels belong to

an organ to calculate the dose for each organ The dose value on each of these structures is sum of self-dose and cross dose This can be used to estimate the risk for vital organs and the effect of treatment in order to extend the life of the patient

Acknowledgements

Many thanks for the support from Mr Nguyen Tan Chau (PET/CT Department-Cho Ray Hospital), Dr Pham Nguyen Thanh Vinh (Ho Chi Minh City University of Education) and

Dr Nguyen Khanh Toan

REFERENCE

1 Helal N (2012) Patient organs dose

calculations in nuclear medicine J Med Signals Sens, 24, 231 – 234.

2 Sergey V Nesterov, Chunlei Han & Maija Mäki et al (2009) Myocardial perfusion

quantitation with 15O-labelled water PET: high reproducibility of the new cardiac analysis

software (Carimas) Eur J Nucl Med Mol Imaging, 36,1594 – 1602.

3 Antoine Rosset, Luca Spadola, Osman Ratib et al (2004) OsiriX: An Open-Source

Software for Navigating in Multidimensional

DICOM Images Journal of Digital Imaging, 17,

205 - 216

4 Alexander K., Pinter C., Fichtinger G (2018) Streamlined open-source gel dosimetry

analysis in 3D slicer Biomedical Physics & Engineering Express, 4, 041 - 045.

5 Yushkevich PA, Piven J, Hazlett HC

và cộng sự (2006) User-guided 3D active

contour segmentation of anatomical structures: Significantly improved efficiency and reliability

Neuroimage, 31, 1116 – 1128.

6 P Arce, J I Lagares, L Harkness (2014) Gamos: A framework to do Geant4

simulations in different physics fields with an

user-friendly interface Nuc Instr Meth, 735,

304 - 313.

7 Jayong Lee and Hoon Kang, (2010)

Flood Fill Mean Shift: A Robust Segmentation

Algorithm International Journal of Control,

Automation, and Systems, 8, 1313 - 1319.

8 Jukka Arvo, Mika Hirvikorpi, Joonas

Tyystjärvi, (2004) Approximate Soft Shadows

with an Image-Space Flood-Fill Algorithm

EUROGRAPHICS, 23, 271 - 280.

9 W S Snyder, M R Ford, G G Warner

et al (1975) “S”, absorbed dose per unit

cumulated activity for selected radionuclides

and organs J Nucl Med, 18, 7 - 717.

10 Susanne Ziegler, Bjoern W Jakoby, Harald Braun (2015) NEMA image quality

phantom measurements and attenuation correction in integrated PET/MR hybrid

imaging EJNMMI Phys, 2(1), 18.

11 Stabin MG, Sparks RB, Crowe E (2005) OLINDA/EXM: the second-generation

personal computer software for internal dose

assessment in nuclear medicine J Nucl Med,

46(6), 7 - 1023.