Brain Tumor Segmentation in Magnetic Resonance Images using Genetic

Algorithm Clustering and AdaBoost Classiﬁer

Gustavo C. Oliveira

, Renato Varoto

and Alberto Cliquet Jr.

1,2,3

University of São Paulo Interunits Graduate Program in Bioengineering, University of São Paulo, São Carlos - SP, Brazil

Department of Orthopedics and Traumatology, University of Campinas (UNICAMP), Campinas - SP, Brazil

Department of Electrical and Computer Engineering, University of São Paulo, São Carlos - SP, Brazil

Keywords:

Image Segmentation, Glioma, Genetic Algorithm, AdaBoost Classiﬁer.

Abstract:

We present a technique for automatic brain tumor segmentation in magnetic resonance images, combining

a modiﬁed version of a Genetic Algorithm Clustering method with an AdaBoost Classiﬁer. In a group of

42 FLAIR images, segmentations produced by the algorithm were compared to the ground truth information

produced by radiologists. The mean Dice similarity coefﬁcient reached by the algorithm was 70.3%. In

most cases, the AdaBoost classiﬁer increased the quality of the segmentation, improving, on average, the

DSC in about 10%. Our implementation of the Genetic Algorithm Clustering method presents improvements

compared to the original method. The use of a ﬁxed, small number of groups and smaller population allowed

for less computational effort. In addition, adaptive restriction in the initial segmentation was achieved by using

the information of the groups with highest and 2nd-highest mean intensities. By exploring intensity and spatial

information of the pixels, the AdaBoost classiﬁer improved segmentation results.

1 INTRODUCTION

Glioma is a type of brain tumor that originates in

the glial cells, which support and surround neurons

in the brain. Growing within the tissue of the brain

and often mixing with healthy areas, it’s the most

frequent primary brain tumor in adults, representing

33% of all cases. By pressing on the spinal cord or

the brain, gliomas can cause many symptoms, such

as seizures, personality changes, weakness in the face

or limbs, problems with speech, vision loss and dizzi-

ness. In its less aggressive form, known as low-grade

gliomas, patients have a life expectancy of several

years, and in its more aggressive form, known as high-

grade gliomas, patients have a median survival rate of

two years or less. The most common treatment for

gliomas is surgery, which may be followed by radia-

tion therapy and chemotherapy (Menze et al., 2015;

Johns Hopkins Medicine Health Library, 2017).

Diagnosis of glioma tumors involves an analysis

of the patient’s medical history, neurological exams

and scans of the brain – magnetic resonance imaging

and computed tomography. Throughout the treatment

process, imaging protocols are used to follow disease

progression and evaluate the success of the chosen

strategy. Analysis of those images usually relies on

rudimentary quantitative measures or qualitative cri-

teria – such as the largest diameter visible from ax-

ial images of the lesion or presence of characteristic

hyper-intense tissue (Menze et al., 2015; Johns Hop-

kins Medicine Health Library, 2017).

In this context, the development of computer

aided-diagnosis (CAD) systems that can automati-

cally analyze brain tumor scans and replace the cur-

rent evaluation methods with more reproducible and

accurate measurements could signiﬁcantly improve

the diagnosis, treatment and follow-up processes by

providing standardized criteria for tumor characteri-

zation and time efﬁciency (Menze et al., 2015; Em-

blem et al., 2009). More speciﬁcally, CADs could

perform the segmentation task, separating the differ-

ent tumor tissues from healthy brain tissue, to extract

the patient speciﬁc clinical information, along with

their diagnostic features. In the last years, a great vari-

ety of segmentation methods was proposed, combin-

ing threshold-based, model-based, region-based and

pixel classiﬁcation techniques (Gordillo et al., 2013).

Comparing these methods, however, is problematic,

since they are usually validated with different perfor-

mance metrics, on small and private datasets and us-

ing different imaging modalities (Menze et al., 2015).

To overcome these difﬁculties, the Multimodal

Oliveira, G., Varoto, R. and Jr., A.

Brain Tumor Segmentation in Magnetic Resonance Images using Genetic Algorithm Clustering and AdaBoost Classiﬁer.

DOI: 10.5220/0006534900770082

In Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2018) - Volume 2: BIOIMAGING, pages 77-82

ISBN: 978-989-758-278-3

Figure 1: Flowchart of the algorithm.

Brain Tumor Image Segmentation (BRATS) chal-

lenge was organized. In this challenge, 20 state-of-

art algorithms were applied to the same dataset and

evaluated in three different tasks: segmentation of

the “whole” tumor (including edema, non-enhancing

solid core, active core and non-solid core), segmen-

tation of the tumor “core” (including all classes ex-

cept edema), and segmentation of the “active tumor”

(containing only the “active core”). It was found that

no single method ranked in the top ﬁve for all tasks -

different algorithms performed best for different sub-

regions. The development of algorithms for brain tu-

mor segmentation continues to attract interest of the

research community (Menze et al., 2015).

Kole and Halder proposed an automatic method

for brain tumor detection and isolation of tumor cells

from magnetic resonance images (MRI) using a ge-

netic algorithm-based clustering method (Kole and

Halder, 2012). In this paper, an extension of their

technique is presented, combining their clustering

method with morphological operation, ﬁlters and an

AdaBoost classiﬁer.

2 MATERIALS AND METHODS

2.1 Algorithm

The ﬂowchart presented in Figure 1 gives an overview

of the proposed technique, which operates as fol-

lows. First, pixels in the image are normalized and

a 3x3x3 mean ﬁlter is applied. Non-background pix-

els are then selected and grouped into clusters ac-

cording to their intensity values using the genetic

algorithm (GA)-based clustering method (Kole and

Halder, 2012).

Let K represent the number of clusters that the

pixels of the image will be divided into. K ranges

from 2 to 5 - i.e. pixels will be divided into 2, 3, 4 and

5 clusters. In the GA-based clustering method, each

chromosome of the population consists of K numbers,

intensity values which represent the centers of each

cluster. At each generation, a ﬁtness function is com-

puted in order to ﬁnd and select the center values that

best divide the pixels into K clusters by minimizing

intra-cluster spread. The next generation is then cre-

ated using mutation and cross-over processes. This

is repeated through several iterations. Then K is in-

creased, new chromosomes are generated and the evo-

lutionary process is repeated. This continues until K

reaches its ﬁnal value. Finally, pixels are grouped us-

ing the most appropriate number of clusters, which is

the one that minimizes the validity index (Kole and

Halder, 2012). Overall, the method can be summa-

rized in the following steps:

1. For each number of groups K, repeat steps 2 to 7.

2. Generate initial values of each chromosome of the

population.

3. Calculate the ﬁtness value of each chromosome.

4. Preserve the chromosome with highest ﬁtness

value found so far.

5. Select chromosomes using the Roulette Wheel al-

gorithm.

6. Create chromosomes of the next generation

through mutation and cross-over.

7. Until termination condition is reached, repeat

steps 3 to 6.

8. Compute the clustering validity index for the

ﬁttest chromosome of all values of K.

9. Cluster pixels using the value of K that minimizes

the validity index.

For more details, please refer to the original article

(Kole and Halder, 2012).

BIOIMAGING 2018 - 5th International Conference on Bioimaging

After ﬁnding the clusters, thresholding is per-

formed to obtain a binary image. Selection of the

threshold value is based on the clusters: the ﬁrst

choice is to use the minimum value of the cluster

with 2nd highest mean intensity. Pixels with inten-

sity value above the threshold are marked as tumor

(i.e. they’re assigned the value “1”), whereas pixels

with intensity value below the threshold are marked

as background (i.e. they’re assigned the value “0”). If

the thresholding operation results in a segmentation

that selects more than 15% of the area of the brain as

tumor, then the ﬁrst binary image is abandoned and

the minimum value of the cluster with highest mean

intensity is chosen as the threshold value. The thresh-

old operation is repeated, producing a new binary im-

age.

Once the binary image is produced, the program

selects the largest connected component and applies

to it a neighborhood ﬁlter. This ﬁlter computes, for

each pixel, the number of tumor pixels that are 8-

connected to it, i.e. the number of tumor pixels that

touches one of the edges or corners of the pixel. If

this number is 4 or more, then the current pixel is in-

cluded. If it’s 2 or less, the current pixel is excluded

or not included in the tumor area (Gibbs et al., 1996;

Sonka et al., 2014). Following the ﬁltering step, holes

in the segmented area are removed by applying the

morphological closing operation iteratively, using a

small disk as the structuring element.

Finally, an AdaBoost classiﬁer is used to improve

the segmentation. The AdaBoost algorithm is a tech-

nique used to create a strong, accurate classiﬁer by

combining weak classiﬁers, assumed to be better than

random guessing in correctly classifying the data. For

a training set of multidimensional data points, a clas-

siﬁer will assign to each data point a label, either +1

or -1. An exponential error function is used to rank

all the weak classiﬁers based on the number of cor-

rect and incorrect classiﬁcations. AdaBoost proceeds

by systematically extracting one classiﬁer of the pool

in each of the iterations, by focusing on the ones that

can help with the misclassiﬁed data points. After ex-

tracting a weak classiﬁer, AdaBoost assigns a weight

to it. The stronger classiﬁer is then given by the group

of extracted weak classiﬁers combined with their as-

signed weights (Freund and Schapire, 1995; Schapire,

1999; Alpaydin, 2014).

In our method, the AdaBoost algorithm classiﬁes

pixels according to three features: intensity value and

coordinates “x” and “y”, which determine the pixel

position in the slice. Preliminary class information,

used as training data, is given by the binary image

produced by the previous steps.

The AdaBoost classiﬁer is applied one slice at a

time, from bottom to top. At each slice, the mean

pixel intensity and standard deviation for healthy

tissue and for tumor tissue are calculated. Non-

background pixels are selected, and data features for

each pixel are extracted. Since AdaBoost seems es-

pecially sensible to noise (Schapire, 1999; Alpay-

din, 2014), pixels that present intensity values that

are more than one standard deviation either above or

below the mean are considered outliers and are ex-

cluded. The selected pixels are then used to train

the classiﬁer, building a model that will be used to

classify pixels in the slice immediately above. This

process continues until there are no more slices to be

classiﬁed. The largest connected component is then

selected, representing the ﬁnal result of the segmenta-

tion process.

In summary, our method uses as basis for an initial

segmentation the GA-based clustering method (Kole

and Halder, 2012). Then, it reﬁnes the segmentation

by applying the neighborhood ﬁlter, the morphologi-

cal closing operation and the AdaBoost classiﬁer.

The algorithm was implemented using MATLAB,

from The Mathworks, Inc. Tests using MRI images

were performed on a Windows 10 PC, 8 GB RAM,

Intel(R) Core(TM) i7-5500U CPU @ 2.40 GHz.

2.2 Image Dataset and Evaluation

Metric

The proposed method was used to perform “whole”

tumor segmentation of low-grade glioma tumors on

tridimensional T2-weighted FLAIR images. These

images, as well as the ground truth information, were

extracted from the 2015 Multimodal Brain Tumor

Image Segmentation Benchmark (BRATS) challenge

database, which is the largest public dataset of its

type, containing a great variety of cases. All of them

were preprocessed in order to homogenize the data

and remove the skulls, guaranteeing anonymization

of the patients (Menze et al., 2015). Two images from

the original database were excluded since the assump-

tion that the largest component represents the tumor

did not hold for them, resulting in a total of 42 test

cases.

Ground truth information was constructed based

on manual annotations performed by a team of trained

radiologists (Menze et al., 2015). Segmentation re-

sults were compared to the ground truth information

using the Dice similarity coefﬁcient (DSC). The DSC

is based on the computation of the area of overlap be-

tween segmented region and ground truth, and it is

considered a very attractive metric because of its sim-

plicity, being widely used for evaluation of segmen-

tation algorithms. It is calculated using the following

Brain Tumor Segmentation in Magnetic Resonance Images using Genetic Algorithm Clustering and AdaBoost Classiﬁer

Table 1: Test cases and the respective DSC values obtained.

Case DSC - Close DSC - Final Case DSC - Close DSC - Final

pat101 0.724426 0.814401 pat325 0.155660 0.697677

pat103 0.372215 0.796679 pat330 0.203043 0.562035

pat109 0.853468 0.811614 pat346 0.447613 0.537141

pat130 0.793343 0.419234 pat351 0.538680 0.743219

pat141 0.857224 0.807989 pat387 0.029067 0.132889

pat152 0.912758 0.735904 pat393 0.461802 0.793748

pat175 0.425288 0.930112 pat402 0.879448 0.880263

pat177 0.095312 0.125464 pat410 0.038867 0.038867

pat202 0.468797 0.473320 pat413 0.586182 0.439421

pat241 0.551048 0.926315 pat420 0.430172 0.716093

pat249 0.732158 0.821051 pat428 0.602043 0.774421

pat254 0.800027 0.931246 pat442 0.141631 0.361035

pat255 0.551579 0.860707 pat449 0.863212 0.831586

pat261 0.704085 0.780075 pat451 0.718508 0.693981

pat266 0.688977 0.791121 pat462 0.590710 0.908526

pat276 0.742505 0.576310 pat466 0.938959 0.950347

pat282 0.668736 0.801328 pat470 0.796363 0.832161

pat298 0.898178 0.769875 pat480 0.495931 0.863931

pat299 0.845593 0.832962 pat483 0.841977 0.912231

pat307 0.458549 0.695856 pat490 0.601234 0.710943

pat312 0.613672 0.703324 pat493 0.856309 0.747976

Average Close DSC 0.594651 Average Final DSC 0.703175

equation:

DSC =

2 × |T

∩ P

| + |P

(1)

where T ∈ {0, 1} and P ∈ {0, 1} are binary maps

representing the ground truth and the algorithm’s seg-

mentation, respectively; T

and P

represent the pixels

where T = 1 and P = 1, respectively; ∩ is the logical

AND operator and |T

| and |P

| represent the size of

the sets T

and P

- the number of pixels belonging to

them (Sonka et al., 2014; Menze et al., 2015).

3 RESULTS

Table 1 presents the DSC values obtained for each test

case by comparing the segmentations produced by the

algorithm to the ground truth information. The “DSC

– Close” columns present the DSC values for the seg-

mentation produced by the steps before the AdaBoost

classiﬁer, while the columns “DSC – Final” present

the DSC values for the ﬁnal segmentation, after the

AdaBoost classiﬁer and selection of the largest com-

ponent. Cases are identiﬁed by the number present in

the original ﬁles from the BRATS database.

The maximum “DSC – Close” and “DSC – Fi-

nal” values obtained were 93.8% and 95.0% (case

“pat466”), with an average “DSC – Close” of 59.4%

and average “DSC – Final” of 70.3%. Segmentation

examples are shown in ﬁgures 2 and 3. In those ﬁg-

ures, the left side presents the original FLAIR images,

while the right side presents the ﬁnal segmentation re-

sults for those cases. The blue regions represent “true

positives”, while the yellow regions represent “false

positives” and the red regions, “false negatives”. The

mean time necessary to segment each image was 6.22

minutes.

4 DISCUSSION

The average ﬁnal DSC value of 70.3% obtained falls

into the expected interval of performance values for

algorithms in this type of application. In comparison,

the state-of-art algorithms tested during the BRATS

challenge using a slightly bigger dataset achieved

mean DSC between 19% and 81% for “whole tumor”

segmentation of low-grade gliomas, with a theoretical

upper limit of individual algorithmic segmentation of

86% and one fused algorithm, created by combining

four different state-of-art methods, achieving mean

DSC of 68% in the same task (Menze et al., 2015).

For the state-of-art algorithms of the BRATS chal-

lenge, average computation times per case ranged

from few minutes to more than an hour, varying sig-

niﬁcantly between algorithms. While a direct com-

BIOIMAGING 2018 - 5th International Conference on Bioimaging

Figure 2: Segmentation example, case “pat466”, DSC 95%.

Figure 3: Segmentation example, case “pat254”, DSC 93%.

parison with our method is not possible, since the

hardware used was different, it is worth noting the rel-

atively small computational time of our method. Re-

duced run time is a valuable feature that should be

pursued, but without forgetting to take into account

the trade-off between computation time and segmen-

tation quality, which is an important part of the pro-

cess of designing a segmentation algorithm for CAD

systems (Menze et al., 2015).

Our implementation of the genetic algorithm clus-

tering method presents some differences from the

original. Through our tests, it was observed that us-

ing a population of 10 individuals and two to ﬁve

groups is sufﬁcient to separate the pixels for the initial

segmentation. Also, selecting between the minimum

values of the clusters with higher mean intensities as

threshold produces higher DSC values, as it allows

for more ﬂexibility in the thresholding operation. In

comparison, (Kole and Halder, 2012) used 30 indi-

viduals, selected tumor pixels from the cluster with

highest mean intensity, and empirically selected the

maximum number of groups for each image.

The percentage used for comparison between the

tumor and brain area was found empirically. This

comparison, together with the selection of the largest

connected component, is useful in eliminating “false

positives” that are created by tissues in the brain that

present themselves in the original FLAIR image as

high intensity regions, even though they are healthy.

In most cases, the AdaBoost classiﬁer increased

the quality of the segmentation, improving, on aver-

age, the DSC in about 10%. The choice of features

– intensity and position in the slice – takes advan-

tage of the fact that, while different tumor structures

Brain Tumor Segmentation in Magnetic Resonance Images using Genetic Algorithm Clustering and AdaBoost Classiﬁer

may present themselves with different intensity val-

ues, they are formed by contiguous regions of tissue,

and their position in consecutive slices do not deviate

much. However, the internal boundaries between tu-

mor tissues and the external boundaries between tu-

mor and healthy tissue, where tissue intensity may

change abruptly, might be a source of error. Including

adequate texture features, for example, may improve

overall performance.

One limitation to our study is the relatively small

number of images available to evaluate the technique,

although it is common practice in the literature to use

small private datasets to evaluate segmentation meth-

ods (Menze et al., 2015). Using more images would

provide a clearer picture of the proposed algorithm’s

performance and areas for improvement. Another

limitation is that the proposed technique heavily re-

lies on pixel intensity information, which is subject to

inter and intra slice variations caused by inhomogene-

ity in the magnetic resonance imaging ﬁeld (Emblem

et al., 2009).

5 CONCLUSION

In conclusion, the method proposed in this paper com-

bines a genetic algorithm-based clustering method

with ﬁlters, morphological operation and an Ad-

aBoost classiﬁer to automatically isolate the tumor

in magnetic resonance images. For the genetic algo-

rithm, improvements were achieved in comparison to

the original version: use of smaller population and a

ﬁxed, small number of groups to perform the cluster-

ing, which allows for less computational effort. An-

other difference is the strategy that makes use of the

groups with highest and 2nd-highest mean intensi-

ties, allowing for adaptive restriction in the initial seg-

mentation. Additionally, the AdaBoost classiﬁer im-

proved segmentation results by taking advantage of

both spatial and intensity information.

Future work may focus on improving the accu-

racy of this technique, by adapting it to evaluate and

include information from other magnetic resonance

modalities, such as T1-weighted. Also, the algorithm

can be further developed by adding methods to ana-

lyze texture features and better tuning of its numerical

parameters, such as the percentage of area for com-

parison between brain and tumor tissue and the num-

ber of rounds used to train the AdaBoost classiﬁer.

Another option is to combine it with other machine-

learning techniques, such as support vector machines,

and create a segmentation based on the consensus of

two or more classiﬁcation methods.

ACKNOWLEDGEMENTS

The authors would like to thank the support by grants

from São Paulo Research Foundation (FAPESP),

Brazilian Federal Agency for Support and Evalu-

ation of Graduate Education (Capes) and National

Council for Scientiﬁc and Technological Develop-

ment (CNPq).

REFERENCES

Alpaydin, E. (2014). Introduction to machine learning.

MIT press.

Emblem, K. E., Nedregaard, B., Hald, J. K., Nome, T.,

Due-Tonnessen, P., and Bjornerud, A. (2009). Auto-

matic glioma characterization from dynamic suscep-

tibility contrast imaging: Brain tumor segmentation

using knowledge-based fuzzy clustering. Journal of

Magnetic Resonance Imaging, 30(1):1–10.

Freund, Y. and Schapire, R. E. (1995). A desicion-theoretic

generalization of on-line learning and an application

to boosting. In European conference on computa-

tional learning theory, pages 23–37. Springer.

Gibbs, P., Buckley, D. L., Blackband, S. J., and Horsman,

A. (1996). Tumour volume determination from mr

images by morphological segmentation. Physics in

medicine and biology, 41(11):2437–2446.

Gordillo, N., Montseny, E., and Sobrevilla, P. (2013). State

of the art survey on mri brain tumor segmentation.

Magnetic resonance imaging, 31(8):1426–1438.

Johns Hopkins Medicine Health Library (2017). Gliomas.

[online] Hopkinsmedicine.org. Available at: https:

//goo.gl/8Ggjp8. [Accessed 15 Jul. 2017].

Kole, D. K. and Halder, A. (2012). Automatic brain tu-

mor detection and isolation of tumor cells from mri

images. International Journal of Computer Applica-

tions, 39(16):26–30.

Menze, B. H., Jakab, A., Bauer, S., Kalpathy-Cramer, J.,

Farahani, K., Kirby, J., Burren, Y., Porz, N., Slot-

boom, J., Wiest, R., et al. (2015). The multimodal

brain tumor image segmentation benchmark (brats).

IEEE transactions on medical imaging, 34(10):1993–

2024.

Schapire, R. E. (1999). A brief introduction to boosting. In

Ijcai, volume 99, pages 1401–1406.

Sonka, M., Hlavac, V., and Boyle, R. (2014). Image

processing, analysis, and machine vision. Cengage

Learning.

BIOIMAGING 2018 - 5th International Conference on Bioimaging