Accepted Manuscript

Big data analysis for brain tumor detection: Deep convolutional neural
networks

Javeria Amin, Muhammad Sharif, Mussarat Yasmin, Steven Lawrence Fernandes

PII: S0167-739X(17)32229-X
DOI: https://doi.org/10.1016/j.future.2018.04.065
Reference: FUTURE 4143

To appear in: Future Generation Computer Systems

Received date : 29 September 2017

Revised date : 14 April 2018
Accepted date : 22 April 2018

Please cite this article as: J. Amin, M. Sharif, M. Yasmin, S.L. Fernandes, Big data analysis for
brain tumor detection: Deep convolutional neural networks, Future Generation Computer Systems
(2018), https://doi.org/10.1016/j.future.2018.04.065

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form.
Please note that during the production process errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.
 Big data analysis for brain tumor detection:
Deep convolutional neural networks
Javeria Amina, Muhammad Sharifa, Mussarat Yasmin*a, Steven Lawrence Fernandesb
a
Department of Computer Science, COMSATS Institute of Information Technology, Wah Cantt, Pakistan
b
Department of Electronics and Communication Engineering, Sahyadri College of Engineering & Management,
Mangaluru, India

javeriacomsat@yahoo.com, muhammadsharifmalik@yahoo.com, mussarat@ciitwah.edu.pk,

steven.ec@sahyadri.edu.in

*Corresponding Author: Mussarat Yasmin (email: mussarat@ciitwah.edu.pk)

Abstract

Brain tumor detection is an active area of research in brain image processing. In this work, a methodology is

proposed to segment and classify the brain tumor using magnetic resonance images (MRI).Deep Neural Networks

(DNN) based architecture is employed for tumor segmentation. In the proposed model, 07 layers are used for

classification that consist of 03 convolutional, 03 ReLU and a softmax layer. First the input MR image is divided

into multiple patches and then the center pixel value of each patch is supplied to the DNN. DNN assign labels

according to center pixels and perform segmentation. Extensive experiments are performed using eight large scale

benchmark datasets including BRATS 2012 (image dataset and synthetic dataset), 2013 (image dataset and synthetic

dataset), 2014, 2015 and ISLES (Ischemic stroke lesion segmentation) 2015 and 2017. The results are validated on

accuracy (ACC), sensitivity (SE), specificity (SP), Dice Similarity Coefficient (DSC), precision, false positive rate

(FPR), true positive rate (TPR) and Jaccard similarity index (JSI) respectively.

Keywords: Random Forests; Segmentation; Patches; Filters; Tissues

1. Introduction

One of the most dreadful kinds of tumors is known as malignant tumors. In adults, Gliomas and lymphomas affect

almost eighty percent cases of malignant tumor [1]. Gliomas include subsets of primary tumor which extent from

low-grade to heterogeneous tumors (more infiltrative malignant tumors). They have maximum prevalence with very

high mortality rate. They can be graded into High Grade Glioma (HGG) and Low Grade Glioma (LGG). HGG is

low infiltrative and aggressive as compared to the LGG. Patients of HGG usually do not survive greater than

1 
 
fourteen months after the detection process. Existing HGG and LGG treatments involve radiotherapy and

chemotherapy [2]. Ischemic stroke is the cerebro vascular infection and common reason of disability and death

worldwide. The affected brain region (stroke lesion) undergoes many stages of the disease categorized as sub-

acute (24h-2w), chronic (>2w) and acute (0-24h) [3].

Segmentation [4] and subsequent quantitative lesions assessment in clinical images provides valuable data for the

evaluation of brain pathologies that are vital for treatment planning methods, disease monitoring prediction and

progression of patient results. Moreover, exact injuries locations relate to specific deficits depending on affected

structure of the brain [5]. The functional deficits produced from stroke lesion are related with damage volume to

specific brain parts [6]. Finally, pathology delineation in an accurate way is most vital step in brain tumor cases in

which estimate of tumor volume of subcomponents region is required for further treatment planning [7]. Correct

segmentation of the lesion region in multi dimensional images is more difficult and challenging work. The lesion

appears in a heterogeneous way such as more variation in size, location, frequency and shape makes it more difficult

to formulate efficient segmentation steps. It’s extremely non-trivial to explain contusions, edema and hemorrhage’s

in subcomponents of brain tumor such as the necrotic core and proliferated cells [8]. The arguably more accurate

segmentation outcomes can be achieved by the manual explanation through human experts that is more time-

consuming, expensive and tedious task. Moreover, it is totally impractical in case of more studies which introduces

additional inter observer variations [9]. More efficient automated method for tumor extraction is a major aim in

computing medical images that provides reproducible, objective and scalable methods for quantitative evaluation of

brain tumor. MS and stroke lesions have same hyper-intense appearance in FLAIR and other white matter lesions

(WML) sequences. It is commonly hard to achieve statistically prior information related to lesion appearance and

shape [10].

Several supervised methods are used for brain lesions segmentation such as Random Forests classifier (RFC),

Intensity based features and generative Gaussian Mixture Model (GMM) [11]. Contextual and morphological [12]

features are used for the detection of different types of brain lesions. Markov Random Field (MRF) [10] is used for

brain lesion segmentation. The above mentioned methods are used with hand crafted feature extraction method but

the problems with hand crafted method are computationally intensive as compared to deep learning methods [13].

At the same time, deep learning methods are more powerful as compared to supervised methods with the great

ability of model to learn more discriminated features for task on hand. These features perform better as well as they

2 
 
are pre-defined and hand-crafted feature sets [14]. Convolutional Neural Networks (CNN) can be used to analyze

the medical imaging problematic tasks to achieve better results. Firstly 2D-CNN is used for segmenting neural

membranes with the support of GPU.3D brain segmentation is obtained via processing the 2D slice separately [15].

Despite the simple architecture, better results are achieved by using these techniques indicating CNN potential.

Large variety of parameters, more computational power and significant memory requirements are needed in fully

3D-CNN model [16]. 2D patches are extracted from multi scale images and further combined into single 3D patches

to avoid the fully 3D-CNN networks. Major reason is discouraging the 3D-CNN usage because it has slow inference

due to computationally more expensive. Hence, classifier biases into rare classes might outcome in over-

segmentation. A CNN model is designed to train samples through distribution of classes that are close to actual class

but over-segmented pixels lead towards incorrect classification in the first phase [17]. Second training phase is also

presented by [14] in which patches on discrimination layer are uniformly extracted from input image. Two phases of

training structure might be prone to over fitting and also more sensitive to first classifier stage. Then dense training

method is used for network training [18]. This method introduced imbalance class label that is similar to uniform

sampling. Weight cost function is used to overcome this problem. Manual adjustment of network sensitivity is

provided but it becomes more difficult to handle multiclass problems by using this method [19].

The overall article organization is as follows: Section II defines related work. Detailed presented approach steps are

mentioned in Section III. DNN outcomes are described in Section IV. Conclusion of this research work is illustrated

in Section V.

2. Related work

In brain tumor cases, atlas can be estimated at the time of segmentation due to the variable location and shape of neo

plasms. Tumor mass effect can be measured by the lesion growth models. The voxels neighborhood provides helpful

information used for obtaining smoother segmentations results by the Markov Random Fields (MRF) [20]. MRF

method is used for brain tumor segmentation. Generative models well generalize hidden data [8] with some

limitations at the training stage. These techniques can learn the pattern of brain tumor without utilizing any specific

model. These types of methods usually consider identical and independent voxels distribution through context

features information. Due to this reason, some small or isolated clusters of voxels may be discriminated mistakenly

in the incorrect class, sometimes in anatomically and physiological improbable locations. To avoid these issues,

many researchers included neighborhood information via embedding probabilistic predictions into a Conditional

3 
 
Random Field (CRF) classifier [21]. Deep CNN models are used to automatically learn hierarchies of complex data

features [22]. CNN can run over patches by using the kernel trick. In tumor segmentation field, recent methods are

utilizing CNN. CNN model is used with max-pooling two convolutional layers and stride with fully-connected (FC)

layer and softmax layer for tumor detection in MR images [23]. 3D [13] and 2D filters [24] are used for the

evaluation of brain tumor. 3D filters can take benefits over the 3D MR nature but computational load is raised.

Two-pathway CNN is used for brain tumor evaluation [14]. Binary CNN is used [24] to detect the complete tumor

region. Then cellular automata are applied to smooth segmentation results before the CNN performs multi-class

discrimination between the sub tumor regions. CNN is trained on extracted patches in each voxel plane and output

of final FC layer, softmax and random forest (RF) are used to train the model [25]. Brain tumor segmentation

process is divided into sub binary tasks and suggested structured predictions on the basis of CNN as a learning

model [26]. Labels of patches are grouped into clusters and CNN predicts input membership on each cluster. Deep

CNN are used for the extraction in MR images. 3×3 kernel size is used to achieve deeper CNN model [27]. Table 1

shows the brief summary of existing methods.

Major contribution of this article is as follows:

1. The proposed DNN model is based on 07 layers hence efficiently segmenting the brain tumor.

2. Input MR image is divided into multiple patches of 4×65×65 and then center pixel label of each patch is

calculated and supplied to the DNN which improves the results of segmentation as well as classification.

3. Proposed model is evaluated on two MICCAI challenges datasets such as ISLES and BRATS.

4. Achieved results are evaluated with recent methods which prove that DNN model performed better than

existing techniques.

Table 1: Summary of existing work

Ref Dataset Methods Results
[8] BRATS2013 Hierarchical majority vote 74%-85% DSC
[21] Contra Cancrum Conditional Random Fields(CRF) 0.84 DSC
[28] ISLES 2015 Ensemble two Deconv Nets 0.67DSC
[23] BRATS 2013 CNN model 83.7±9.4 DSC
[13] Brats 2015, ISLES 3D fully connected conditional random 84.7 DSC
2015 field
[26] BRATS2014 CNNs model 83±13 DSC
[27] BRATS 2013 CNN with small (3x3) filters 0.88 DSC
[14] BRATS 2013 Input Cascade (CNN) 0.81DSC
[29] ISLES 2015 CNN model 69%DSC
[30] BRATS 2012(image) CRF (conditional random fields) 62%DSC

4 
 
3. Propossed CNN arch
hitecture

A patch based
b method is presented foor problem off brain tumor ddetection in M
MR images. Thee input MRI im
mage is first

divided innto N patches.. The center piixel label of eaach patch is thhen estimated by using a traiined CNN moodel. Overall

results are then generatted by combinning the predictions for all ppatches. Due to low resoluttion in third ddimension of

MR imagges, segmentaation is perfoormed on eacch slice from

m different axxial views. Thhe proposed architecture

processess sequentially every 2D slicce in which eaach pixel is rellated through different MR modalities e.gg., diffusion

weighted image (DWI)), fluid attenuaation inversionn recovery (FL

LAIR), spin-sppin relaxation (T2), T1 and T1- contrast

and image modalities liike many CNN

N segmentatioon methods [144]. Proposed D
DNN model iss shown in Figgure 1.

Fig.1.Propossed model (paatch generationn, CNN train m

model, label generation, seggmented imagee)

ning Phase
3.1 Train

The inputt image to prooposed DNN m

model is a muultiple channel MRI image of size 4×2400×240 shown in Figure 2.

The inputt image is firsst divided of4×

×65×65 patchhes. The Inputt X of proposed DNN methhod is accordingly a N×N

2D patch with all MR modalities. The primary buuilding block utilized to devvelop DNN model
m is the coonvolutional

layer. Maany layers whhich form featuures hierarchyy could be staacked on the ttop of one anoother. Each laayer extracts

features ffrom its previoous layer in thhe hierarchy too which it is liinked. Each coonvolutional laayer executes MR images

planes ass an input staack and deliveers some num

mber of featuure maps as aan output. Eacch feature maap could be

topologiccally organizeed in the respponse maps of

o specific noon-linear spattial feature eextraction callled learning

parameterrs which are iidentically appplied to everyy spatial neighhborhood inpuut planes in thhe sliding winddow design.

i convolutionnal layer process, the singgle plane of iinput images is related to various MR modalities.
Initially in

Followingg the layers, thhe informationn planes normally consist off feature mapss of the precedding layer.

5 
 
 Figg.2.Procedure of proposed D
DNN training m
modal

In the pproposed trainning model, 55×5 filterωis applied withh convolutionnal layers ouutput size N m 1

N m 1 . Non-lineearity pre-inpuut is computedd x ,

in convolutional laayers and filter component iis needed to

be added,, weighted on the previous llayers. This is given in Eq. ((1) and Eq. (2)) below.

x , ω y 1

y ,
σ x , 2

In this opperation, each feature map taakes maximum

m neuron (featture) value in 3×3 sub winddows. Max-poooling makes

position iinvariance oveer larger neighhborhood areaas and down specimens

s the information ppicture by a faactor of 3×3

along eveery direction [331].

Max-poolling prompts quicker convvergence rate by choosing predominant invariant feaatures which enhance
e the

generalizaation executioon. The outpuut feature mappping process is performed at the convollutional layer as an input

from the previous convvolutional layeers. With the neural networrk perception, feature mapss are related too the hidden

layers off neurons. Eacch coordinate of feature m

map is linked to a single nneuron in which receptive field size is

accordingg to the size off kernels. A keernel value deenotes the connnections by w
weights among the neurons laayers. In the

process oof learning, eaach kernel is adjusted to a distinct spaatial frequencyy orientation after which iit scales the

training ddata statistics.. Finally, preddiction is perrformed on thhe basis of testing and traiining labels. E
Each kernel

detects tiissues on thee basis of seggmentation laabels. In the proposed DN

NN model, prre-training prooperties are

6 
 
connected to the layered structure of neurons, such receptive field values and spatial arrangements are called hyper

parameters that are mentioned in Table 2 and Table 3.

Table2: Hyper parameters of presented DNN model

Momentum 0.9
L2 Regularization 0.004
Max-Epochs 2
Mini-Batch Size 128

Table3: Hyper parameters regarding Learning rate of proposed DNN model

Learning Rate Value
Initial 0.001
Schedule Piecewise
Drop Factor 0.1
Drop Period 8

3.2 Testing Phase

The full MR image is fed as an input. Then it is converted into 4×65×65 patches. Furthermore, output layer which is

the nature of convolutional, permits us to more computations time at the testing phase. Therefore, convolutions

perform at all the layers to obtain all labels (at the center pixel value) probabilities p y |x for the whole MR image.

Then softmax layer is applied due to its non-linear properties to normalize results of convolutions kernel. Let d

denote vector value at a spatial position to calculate softmax d where z ∑ exp d is a constant

normalization and y denotes label of the input image such that it interprets every spatial convolutional output layer

position as a distribution architecturep y |x in which y , gives the label position at i, j. The proposed model assigns

label to each pixel with the highest probability. Fig 3 shows filter results from first convolutional layer of proposed

DNN architecture.

7 
 
Fig.3.Filtter results on tthe benchmarkk datasets (a) BRATS 20133 (b) BRATS 2014 (c) ISL
LES 2015 (d) ISLES
I 2017
(e) BRAT TS 2012 (f) syynthetic 2013(gg) synthetic 20012 (h) BRAT
TS 2015

4. Results and experim

mentation

The proposed architectture is evaluatted on eight bbenchmark daatasets such ass BRATS 20112, 2013, 20144, 2015 and

ISLES (IIschemic strokke lesion segm

mentation) 20015 and 20177. The descripption of propoosed training and testing

subjects iin benchmark datasets is desscribed as follows:

TS 2015 dataset, total 384 ssubjects are used in which 220 HGG witth 54 LGG suubjects are in the training
In BRAT

phase andd 110 both (L

LGG+HGG) subjects
s t proposed model [8]. BR
are appplied in testiing phase of the RATS 2012

image dattaset contains multimodal (T

T1, T2, T1-coontrast, Flair) MR
M images inn which total 80
8 input subjeccts are used.

25 LGG and 25 HGG

G are used in the training pphase and 20 HGG and 10 LGG testingg subjects are used in the

proposed model [32]. IIn BRATS 2014 dataset, 3000 subjects aree used in whicch 200 trainingg and 100 testting subjects

8 
 
are taken in the proposed model [46]]. BRATS20133 image dataset consists of 30 input subjeects in which 220HGG and

10 LGG ssubjects are taaken in traininng stage and 10

1 both (LGG and HGG) teesting subjectss are used in thhe proposed

model [477]. BRATS2013 synthetic dataset

d consistts of 50 gliomaa subjects in which
w 25 both (LGG+HGG)) are used in

training aand remainingg 25 testing suubjects are utiilized in the proposed modeel [47]. BRAT
TS 2012 synthhetic dataset

contains 50 subjects inn which half ssubjects are used in trainingg and half aree used in testiing stage of thhe proposed

model [488]. ISLES 20115 dataset connsists of 64 Suub-Acute Strokke Lesion Seggmentation (SIISS) subjects in which 28

training aand 36 testingg subjects aree used in the proposed moddel. It containns four MRI modalities suuch as DWI,

FLAIR, T
T2, T1 and T11-contrast [333]. ISLES 2017 dataset connsists of 75 inpput subjects; 443 training annd 32 stroke

testing suubjects are useed in the proposed model [334]. The proposed segmenttation results oon all MRI moodalities are

shown in Fig 4. The overall experim

mentation is performed on C
Core i7 3.4 G
GHz CPU withh 32GB RAM
M and Nvidia

K40 GPU
U running on thhe MATLAB toolbox. 

Fig.4.Sammple segmenttation results of the preseented method,, (upper row represents thhe input imaages of five
modalitiees and lower roow depicts thee correspondinng segmentatioon results) (a) DWI (b) Flairr (c) T1-c (d) T2
T (e) T1

4.1 Perfoormance evalu

uation

Completee region of tuumor (includinng IV classes of intra-tumooral region, laabels I (necrossis), II (edemaa), III (non-

enhancingg tumor) andd IV (enhancinng tumor) is validated on eight perform

mance metricss such as SE
E, SP, ACC,

precision,, FPR, FNR, JJSI and DSC [[35]. In Eq. (33) to Eq. (10), true positive (TP) represennts total tumor region, true

negative ((TN) shows aall non-tumor pixels,

p false nnegative (FN) ddescribes tum
mor pixels that proposed moddel does not

classify aand false positiive (FP) givess tumor pixels that are wronngly classifiedd as being tumor. JSI is usedd to measure

9 
 
similarity between the sets. Its values lie among 1 and 0 and greater value shows accurate results of segmentation.

DSC calculates the automatic and manual segmentation defined below.

2TP
DSC 3
FP 2TP FN

TP
SE 4
TP FN

TN
SP 5
TN FP

TP TN
ACC 6
TP TN FP FN
TP
Precision 7  
TP FP

FPR 1 Specificity 8  

FNR 1 Sensitivity 9  

TP
JSI 10
TP FN FP

4.2 Discussion and experimental results

The key components that affect and extract patches on the proposed model performance are analyzed. The presented

method comparison with the previous deep learning techniques is mentioned in this section. Finally outcomes of

proposed technique are reported by experimentation on all modalities of BRATS and ISLES datasets and shown in

Table 4 and Table 5. Proposed method performance is also evaluated on individual MR modalities given in Table 6.

Computational time is also considered regarding each image modality as well as overall average time is mentioned

in Table 7.

Table 4: Proposed method results on all BRATS datasets

Dataset DSC SE JSI FNR FPR Precision SP ACC
BRATS 2012 (image) 98.4% 98.5% 96.0% 0.02 0.02 99.9% 98.4% 98.6%
BRATS 2012(syntehtic image) 100% 100% 100% 0.00 0.00 100% 100% 100%
BRATS 2013 (image) 99.8% 99.7% 99.0% 0.01 0.01 98.9% 99.9% 99.8%
BRATS 2013(syntehtic image) 100% 100% 100% 0.00 0.00 100% 100% 100%
BRATS 2014 92.9% 93.0% 85.1% 0.08 0.07 95.5% 93.2% 93.1%
BRATS 2015 95.0% 95.0% 90.4% 0.05 0.05 97.2% 95.2% 95.1%

Table 5: Proposed method results on ISLES datasets

Dataset DSC SE JSI FNR SP FPR Precision ACC
ISLES 2015 100% 100% 100% 0.00 100% 0.00 100% 100%
ISLES 2017 98.7% 98.7% 96.0% 0.02 98.8% 0.02 99.0% 98.8%

10 
 
 Table 6: Performance on each image modality
Dataset DSC SE JSI FNR FPR Precision SP ACC
Flair 99.8% 99.7% 99.0% 0.01 0.01 98.9% 99.8% 99.9%
DWI 100% 100% 100% 0.00 0.00 100% 100% 100%
T2 98.0% 96.2% 96.0% 0.04 0.00 100% 100% 97.0%
T1-C 95.4% 96.1% 90.4% 0.06 0.16 94.8% 84.6% 93.2%
T1 97.4% 94.9% 94.1% 0.06 0.00 100% 100% 96.1%

Table 7: Proposed method computational time

Modality Computational time
Flair 6.856sec
DWI 0.008 sec
T2 6.865sec
T1-C 6.853 sec
T1 6.932 sec
Average time 5.502 sec

On all MRI modalities, proposed method achieved 98.4% DSC, 98.5% SE, 96.0 JSI, 0.02 FNR, 0.02 FPR, 99.9%

precision, 98.4% SP, 98.65% ACC and 99.8% DSC, 99.7% SE, 99.0% JSI, 0.01FNR, 0.01 FPR, 99.9% SP, 99.8%

ACC on BRATS 2012 and BRATS 2013 (image) datasets. In the same way it obtained 100% results by using all

performance measures in BRATS 2012 and BRATS 2013 (synthetic) datasets. However, BRATS 2015 achieved

better results as compared to BRATS 2014 dataset. Similarly proposed method is tested on ISLES 2015 and 2017

datasets in which ISLES 2015 obtained higher accuracy as compared to ISLES 2017 dataset.

Proposed method performance is also validated on individual MRI modalities on all benchmark datasets. In this

scenario, it is observed that Flair and DWI achieve greater outcomes as compared to other modalities. It obtained

99.8% DSC, 99.7% SE, 99.0% JSI, 0.01 FNR, 0.01 FPR, 98.9% precision, 99.8% SP, 99.9% ACC and 100% DSC,

100% SE, 100 JSI, 0.00 FNR, 0.00 FPR, 100% precision, 100% SP, 100% ACC on Flair and DWI modalities. In the

same case of other modalities such as T1, T1-contrast and T2, T2 achieves better results as compared to T1 and T1-

contrast. T2 obtains 98.0% DSC, 96.2% SE, 96.0% JSI, 0.04 FNR, 0.00 FPR, 100% SP and 97.0% ACC

respectively. Moreover, after the performance evaluation on each modality, it is observed that precision rate is

increased on 40 training epochs in T1, T1-contrast and T2 modalities but Flair and DWI achieve better results in less

than 40 training epochs. The proposed methodology achieved better results in less processing time as compared to

the recently existing architecture. Computational time is also compared between each modality and it is observed

that Flair, DWI, T1, T1-contrast, T2 take 6.856 sec, 0.008 sec, 6.932 sec, 6.853 sec and 6.865 sec in segmentation

process respectively. The average processing time of the proposed DNN model is 5.502 sec which proves that the

DNN model is light in nature. Comparison of the presented approach performance with the existing techniques is

11 
 
illustrated in Tables 8, 9, 10, 11, 12 and 13. In terms of computational time, proposed model comparison with the

existing models is given in Table 14.

Table 8: Proposed method results on ISLES-SISS 2015 dataset

Method DSC SE
Proposed 100% 100%
Chen et al.[28] 67.0% -
Haeck et al.[36] 78% 80%
Larochelle et al.[29] 69% 67%
McKinley et al.[37] 85% -
Mahmood et al.[38] - 50 %

Table 9: Proposed method results on BRATS 2013 (synthetic) dataset

Method DSC
Proposed 100%
Abbasi et al.[39] 93%
Cordier et al.[40] 84%

Table 10: Proposed method results on BRATS 2015 dataset

Method DSC SE SP ACC
Proposed 95% 95% 95.2% 95.1%
Dong et al.[41] 86% - - -
Pereira et al.[27] 78% - - -
Havaei et al.[14] 79% - - -
Kamnitsas et al.[13] 90% 90.4% - -

Table 11: Proposed method results on BRATS 2014 dataset

Method DSC SE SP ACC
Proposed 92.9% 93% 93.2% 93.1%
Reza et al.[42] 89.6% 84% - 82.2%

Table 12: Proposed method results on BRATS 2012 (image) dataset

Method DSC
Proposed 98.4%
Wu et al.[30] 62%
Bauer et al. [32] 73%
Huang et al. [43] 75%

Table 13: Proposed method results on BRATS 2013 (image) dataset

Method DSC SE SP ACC
Proposed 99.8% 99.7% 99.9% 99.8%
Zikic et al.[44] 83.7% - - -
Reza et al.[42] 90.9% - - 86.7%
Havaei et al.[14] 88% 89% 87% -
Pereira et al.[27] 88% - - -
Goetz et al.[45] 83% - - -
Huang et al.[43] 88% - - -

12 
 
 Table 14: Computational time comparison of proposed method with existing methods
Method Computational time for the prediction of per brain image
Tustison’s method[8] 100 minutes
Input Cascade CNN[14] 3 minutes
Two Path CNN[14] 25 seconds
Proposed DNN model 5.502 sec

Presented approach performance is evaluated with seventeen previous techniques such as [28], [36], [29], [37], [38],

[39], [40], [41], [27], [14], [13], [42], [30], [32], [43], [44] and [45]. Two Deconv Networks are ensemble in which

one is EDD and second one is CNN for brain tumor detection. It is tested on 741 stroke lesion images. It obtained

94% DSC [28]. Expectation maximization method is used for stroke lesion detection. It achieved 78% DSC and

80% SE [36]. CNN model is taken into account for stroke lesion detection [29]. A Decision Forest method with

threshold finding approch and RFC is used for stroke lesion segmentation [37, 38]. Otsu and Random Forest method

is used for glioma detection. This method is tested on BRATS 2013 synthetic dataset. It achieved 93% DSC [39].

Patch based segmentation method is used for brain lesion detection [40]. U-Net Based Fully CNN model is utilized

for the detection of brain lesion and it achieved 86% DSC on BRATS 2015 dataset [41]. A CNN model is tested for

brain tumor detection [27]. Input Cascade model (CNN) model is tested on BRATS 2013 image dataset for detecting

brain lesion [14]. 3D fully connected conditional random field is used for the detection of brain lesion. It obtained

90% DSC and 90.4% SE respectively [13]. Texture features are used for brain tumor detection. This approach

achieved 89.6% DSC, 84% SE and 82.2% ACC on BRATS 2014 dataset [42]. Conditional random fields (CRF)

method with the pixel-pairwise affinity and superpixel-level features are used for glioma detection [30]. Hierarchical

regularization and classification method is tested on BRATS 2012 image dataset. It attined 76% DSC [32]. Local

independent projection-based classification (LIPC) technique is utilized to classify the class labels. It is tested on

BRATS 2013 image dataset and obtained 88% DSC [43]. CNN architecture [44] and extremely randomized trees

[45] is used for glioma detection. It obtained 83.7% and 83.0% DSC on BRATS 2013 image dataset.

Through observing the above mentioned experimental results, the performance of proposed technique is better as

compared to previous approaches which demonstrate applicability of proposed model.

13 
 
5. Conclusion

In this work, DNN based architecture is proposed for brain tumor detection. Proposed model is evaluated on eight

challenges datasets and five MRI modalities such as Flair, DWI, T2, T1 and T1-contrast respectively. The achieved

results are 99.8% DSC on Flair, 100% results on DWI, 98.0% on T2, 97.4% on T1 and 95.4% on T1-contrast

modalities. The proposed model is validated on a number of datasets and performance measures while the existing

techniques are not evaluated on such number of datasets and performance measures. This authenticates the

consistency of proposed model because it equally performs well on all datasets and performance measures. These

results are compared with existing methods which show that proposed model performs better in terms of accuracy

and time because average processing time of the proposed CNN model is 5.502 sec. This model practically can be

used for brain lesion detection at an early stage.

References

[1]  J. Amin, M. Sharif, M. Yasmin, and S. L. Fernandes, “A distinctive approach in brain tumor 
detection and classification using MRI,” Pattern Recognition Letters, 2017. 
[2]  V. Rajinikanth, S. L. Fernandes, B. Bhushan, and N. R. Sunder, "Segmentation and Analysis of 
Brain Tumor Using Tsallis Entropy and Regularised Level Set." pp. 313‐321. 
[3]  V. Rajinikanth, and S. C. Satapathy, “Segmentation of Ischemic Stroke Lesion in Brain MRI Based 
on Social Group Optimization and Fuzzy‐Tsallis Entropy,” Arabian Journal for Science and 
Engineering, pp. 1‐14, 2018. 
[4]  V. Rajinikanth, S. C. Satapathy, S. L. Fernandes, and S. Nachiappan, “Entropy based segmentation 
of tumor from brain MR images–a study with teaching learning based optimization,” Pattern 
Recognition Letters, vol. 94, pp. 87‐95, 2017. 
[5]  D. J. Sharp, C. F. Beckmann, R. Greenwood, K. M. Kinnunen, V. Bonnelle, X. De Boissezon, J. H. 
Powell, S. J. Counsell, M. C. Patel, and R. Leech, “Default mode network functional and structural 
connectivity after traumatic brain injury,” Brain, vol. 134, no. 8, pp. 2233‐2247, 2011. 
[6]  L. M. Carey, R. J. Seitz, M. Parsons, C. Levi, S. Farquharson, J.‐D. Tournier, S. Palmer, and A. 
Connelly, “Beyond the lesion: neuroimaging foundations for post‐stroke recovery,” Future 
Neurology, vol. 8, no. 5, pp. 507‐527, 2013. 
[7]  P. Y. Wen, D. R. Macdonald, D. A. Reardon, T. F. Cloughesy, A. G. Sorensen, E. Galanis, J. 
DeGroot, W. Wick, M. R. Gilbert, and A. B. Lassman, “Updated response assessment criteria for 
high‐grade gliomas: response assessment in neuro‐oncology working group,” Journal of clinical 
oncology, vol. 28, no. 11, pp. 1963‐1972, 2010. 
[8]  B. H. Menze, A. Jakab, S. Bauer, J. Kalpathy‐Cramer, K. Farahani, J. Kirby, Y. Burren, N. Porz, J. 
Slotboom, and R. Wiest, “The multimodal brain tumor image segmentation benchmark 
(BRATS),” IEEE transactions on medical imaging, vol. 34, no. 10, pp. 1993‐2024, 2015. 
[9]  E. L. Yuh, S. R. Cooper, A. R. Ferguson, and G. T. Manley, “Quantitative CT improves outcome 
prediction in acute traumatic brain injury,” Journal of neurotrauma, vol. 29, no. 5, pp. 735‐746, 
2012. 

14 
 
[10]  J. Mitra, P. Bourgeat, J. Fripp, S. Ghose, S. Rose, O. Salvado, A. Connelly, B. Campbell, S. Palmer, 
and G. Sharma, “Lesion segmentation from multimodal MRI using random forest following 
ischemic stroke,” NeuroImage, vol. 98, pp. 324‐335, 2014. 
[11]  R. Domingues, M. Filippone, P. Michiardi, and J. Zouaoui, “A comparative evaluation of outlier 
detection algorithms: Experiments and analyses,” Pattern Recognition, vol. 74, pp. 406‐421, 
2018. 
[12]  C. Ledig, R. A. Heckemann, A. Hammers, J. C. Lopez, V. F. Newcombe, A. Makropoulos, J. 
Lötjönen, D. K. Menon, and D. Rueckert, “Robust whole‐brain segmentation: application to 
traumatic brain injury,” Medical image analysis, vol. 21, no. 1, pp. 40‐58, 2015. 
[13]  K. Kamnitsas, C. Ledig, V. F. Newcombe, J. P. Simpson, A. D. Kane, D. K. Menon, D. Rueckert, and 
B. Glocker, “Efficient multi‐scale 3D CNN with fully connected CRF for accurate brain lesion 
segmentation,” Medical image analysis, vol. 36, pp. 61‐78, 2017. 
[14]  M. Havaei, A. Davy, D. Warde‐Farley, A. Biard, A. Courville, Y. Bengio, C. Pal, P.‐M. Jodoin, and H. 
Larochelle, “Brain tumor segmentation with deep neural networks,” Medical image analysis, vol. 
35, pp. 18‐31, 2017. 
[15]  A. de Brébisson, and G. Montana, “Deep neural networks for anatomical brain segmentation,” 
arXiv preprint arXiv:1502.02445, 2015. 
[16]  H. R. Roth, L. Lu, A. Seff, K. M. Cherry, J. Hoffman, S. Wang, J. Liu, E. Turkbey, and R. M. 
Summers, "A new 2.5 D representation for lymph node detection using random sets of deep 
convolutional neural network observations." pp. 520‐527. 
[17]  D. Ciresan, A. Giusti, L. M. Gambardella, and J. Schmidhuber, "Deep neural networks segment 
neuronal membranes in electron microscopy images." pp. 2843‐2851. 
[18]  K. Simonyan, and A. Zisserman, “Very deep convolutional networks for large‐scale image 
recognition,” arXiv preprint arXiv:1409.1556, 2014. 
[19]  O. Ronneberger, P. Fischer, and T. Brox, "U‐net: Convolutional networks for biomedical image 
segmentation." pp. 234‐241. 
[20]  B. H. Menze, K. Van Leemput, D. Lashkari, M.‐A. Weber, N. Ayache, and P. Golland, "A 
generative model for brain tumor segmentation in multi‐modal images." pp. 151‐159. 
[21]  S. Bauer, L.‐P. Nolte, and M. Reyes, "Fully automatic segmentation of brain tumor images using 
support vector machine classification in combination with hierarchical conditional random field 
regularization." pp. 354‐361. 
[22]  Y. Bengio, A. Courville, and P. Vincent, “Representation learning: A review and new 
perspectives,” IEEE transactions on pattern analysis and machine intelligence, vol. 35, no. 8, pp. 
1798‐1828, 2013. 
[23]  D. Zikic, Y. Ioannou, M. Brown, and A. Criminisi, “Segmentation of brain tumor tissues with 
convolutional neural networks. in proc of BRATS‐MICCAI,” 2014. 
[24]  M. Lyksborg, O. Puonti, M. Agn, and R. Larsen, "An ensemble of 2D convolutional neural 
networks for tumor segmentation." pp. 201‐211. 
[25]  V. Rao, M. Sarabi, and A. Jaiswal, “Brain tumor segmentation with deep learning,” MICCAI 
Multimodal Brain Tumor Segmentation Challenge (BraTS), pp. 56‐59, 2015. 
[26]  P. Dvorak, and B. Menze, “Structured prediction with convolutional neural networks for 
multimodal brain tumor segmentation,” Proceeding of the Multimodal Brain Tumor Image 
Segmentation Challenge, pp. 13‐24, 2015. 
[27]  S. Pereira, A. Pinto, V. Alves, and C. A. Silva, “Brain tumor segmentation using convolutional 
neural networks in MRI images,” IEEE transactions on medical imaging, vol. 35, no. 5, pp. 1240‐
1251, 2016. 
[28]  L. Chen, P. Bentley, and D. Rueckert, “Fully Automatic Acute Ischemic Lesion Segmentation in 
DWI Using Convolutional Neural Networks,” NeuroImage: Clinical, 2017. 

15 
 
[29]  H. Larochelle, and P.‐M. Jodoin, "A Convolutional Neural Network Approach to Brain Tumor 
Segmentation." p. 195. 
[30]  W. Wu, A. Y. Chen, L. Zhao, and J. J. Corso, “Brain tumor detection and segmentation in a CRF 
(conditional random fields) framework with pixel‐pairwise affinity and superpixel‐level 
features,” International journal of computer assisted radiology and surgery, vol. 9, no. 2, pp. 
241‐253, 2014. 
[31]  D. Scherer, A. Müller, and S. Behnke, “Evaluation of pooling operations in convolutional 
architectures for object recognition,” Artificial Neural Networks–ICANN 2010, pp. 92‐101, 2010. 
[32]  S. Bauer, T. Fejes, J. Slotboom, R. Wiest, L.‐P. Nolte, and M. Reyes, "Segmentation of brain tumor 
images based on integrated hierarchical classification and regularization." 
[33]  O. Maier, M. Wilms, J. von der Gablentz, U. M. Krämer, T. F. Münte, and H. Handels, “Extra tree 
forests for sub‐acute ischemic stroke lesion segmentation in MR sequences,” Journal of 
neuroscience methods, vol. 240, pp. 89‐100, 2015. 
[34]  M. Descoteaux, L. Maier‐Hein, A. Franz, P. Jannin, D. L. Collins, and S. Duchesne, Medical Image 
Computing and Computer‐Assisted Intervention− MICCAI 2017: Springer, 2007. 
[35]  L. R. Dice, “Measures of the amount of ecologic association between species,” Ecology, vol. 26, 
no. 3, pp. 297‐302, 1945. 
[36]  T. Haeck, F. Maes, and P. Suetens, "ISLES challenge 2015: Automated model‐based 
segmentation of ischemic stroke in MR images." pp. 246‐253. 
[37]  R. McKinley, L. Häni, R. Wiest, and M. Reyes, "Segmenting the Ischemic Penumbra: A Decision 
Forest Approach with Automatic Threshold Finding." pp. 275‐283. 
[38]  Q. Mahmood, and A. Basit, "Automatic Ischemic Stroke Lesion Segmentation in Multi‐spectral 
MRI Images Using Random Forests Classifier." pp. 266‐274. 
[39]  S. Abbasi, and F. Tajeripour, “Detection of brain tumor in 3D MRI images using local binary 
patterns and histogram orientation gradient,” Neurocomputing, vol. 219, pp. 526‐535, 2017. 
[40]  N. Cordier, B. Menze, H. Delingette, and N. Ayache, "Patch‐based segmentation of brain 
tissues." pp. 6‐17. 
[41]  H. Dong, G. Yang, F. Liu, Y. Mo, and Y. Guo, “Automatic Brain Tumor Detection and 
Segmentation Using U‐Net Based Fully Convolutional Networks,” arXiv preprint 
arXiv:1705.03820, 2017. 
[42]  S. M. Reza, R. Mays, and K. M. Iftekharuddin, "Multi‐fractal detrended texture feature for brain 
tumor classification." 
[43]  M. Huang, W. Yang, Y. Wu, J. Jiang, W. Chen, and Q. Feng, “Brain tumor segmentation 
based on local independent projection‐based classification,” IEEE transactions on 
biomedical engineering, vol. 61, no. 10, pp. 2633‐2645, 2014. 
[44]  D. Zikic, Y. Ioannou, M. Brown, and A. Criminisi, “Segmentation of brain tumor tissues with 
convolutional neural networks,” Proceedings MICCAI‐BRATS, pp. 36‐39, 2014. 
[45]  M. Goetz, C. Weber, J. Bloecher, B. Stieltjes, H.‐P. Meinzer, and K. Maier‐Hein, “Extremely 
randomized trees based brain tumor segmentation,” Proceeding of BRATS challenge‐MICCAI, pp. 
006‐011, 2014. 

[46] BRATS2014, https://sites.google.com/site/miccaibrats2014/(accessed last time, 3/28/2018).

[47] BRATS2013,https://www.smir.ch/BRATS/Start2013 (accessed last time, 3/28/2018).
[48] BRATS2012, https://www.smir.ch/BRATS/Start2012/(accessed last time, 3/28/2018).

16 
 
 
Javeria Amin has completed Bachelor in Software Engineering from UET Taxila, Pakistan in 2012 and MS (CS) from 
COMSATS Wah Cantt, Pakistan in 2016. Her area of specialization is Image Processing. She received scholarships 
during her masters. Currently she is a student of PhD in COMSATS Wah Cantt Pakistan. Her research interests are 
Artificial Intelligence and Neural Networks. 

Muhammad Sharif, PhD is Associate Professor at COMSATS, Wah Cantt Pakistan. His area of 
specialization is Artificial Intelligence and Image Processing. He is into teaching field from 1995 to date. 
He has 110 plus research publications in IF, SCI and ISI journals and national and international 
conferences. He has so far supervised 25 MS (CS) thesis. He is currently supervising 5 PhD (CS) students 
and co‐supervisor of 5 others. More than 200 undergraduate students have successfully completed their 
project work under his supervision. His research interests are Image Processing, Computer Networks & 
Security and Algorithms Design and Analysis. 
Mussarat Yasmin, PhD is Assistant Professor at COMSATS, Wah Cantt Pakistan. Her area of 
specialization is Image Processing. She is in education field since 1993. She has so far 30 research 
publications in IF, SCI and ISI journals as well as in national and international conferences. A number of 
undergraduate projects are complete under her supervision. She is currently supervising 5 PhD (CS) 
students. She is gold medallist in MS (CS) from IQRA University, Pakistan. She is getting COMSATS 
research productivity award since 2012. Her research interests include Neural Network, Algorithms 
design and Analysis, Machine Learning and Image processing. 
 

Steven Lawrence Fernandes, PhD is member of Core Research Group, Karnataka Government Research Centre of
Sahyadri College of Engineering and Management, Mangalore, Karnataka. He has received Young Scientist Award
by Vision Group on Science and Technology, Government of Karnataka. He also received grant from The
Institution of Engineers (India), Kolkata for his Research work. His current Ph.D. work, “Match Composite Sketch
with Drone Images”, has received patent notification (Patent Application Number: 2983/CHE/2015) from the
Government of India.
 

17 
 
Javeria Amin

Muhammad Sharif

Mussarat Yasmin

Steven L. Fernnades
 

Research Highlights

 A new light-weight Deep Neural Networks approach for brain tumor segmentation.
 Extensive evaluation of proposed model on eight challenging big datasets.
 Proposed work achieves state-of-the-art accuracy on these benchmark datasets.
 Comparison of presented work with sixteen existing techniques in the same domain.
 Better results by proposed method without incurring heavy computational burden.

19