Final_ppt – Copy-min

I N F O R M AT I O N T E C H N O L O G YE L E C T R I C A L E N G I N E E R I N G A N DFA C U LT Y O F2019 Tumor Segmentation Using CNN j 1/27

Brain Tumor Segmentation Using Convolutional

Neural Networks in MRI Images

IEEE TRANSACTIONS ON MEDICAL IMAGING,VOL.35,NO.5,MAY2016

S

 ergio Pereira, Adriano Pinto, Victor Alves, and Carlos A.

Silva

Student presentation of the Scientifc working class

Otto-von-Guericke-Universit

 at, Magdeburg

Created and presented by: Divya Thomas

12 November,2019

I N F O R M AT I O N T E C H N O L O G YE L E C T R I C A L E N G I N E E R I N G A N DFA C U LT Y O F2019 Tumor Segmentation Using CNN j 2/27

Overview

1. Introduction

2. Proposed methodology

3. Experimental setup

4. Results and Discussions

5. Literature

2019 Tumor Segmentation Using CNN j 3/27

Introduction

I mageS egmentation

Segment an image into meaningful

regions for a particular

application(Menze, Jakab, et al.

2014).

It can be grey level, colour,

texture,depth or motion.

Examples for image

segmentation (Menze, Jakab,

et al. 2014)

2019 Tumor Segmentation Using CNN

j4/27 Introduction Cont.

C onvolutionalN euralN etwork (C N N )

Gives mathematical representation

of how neurons and synapses work

in the human brain.

CNN consists of an input, output

and multiple hidden layers (Menze,

Jakab, et al. 2014). Neural networking (Menze,

Jakab, et al. 2014)

2019 Tumor Segmentation Using CNN

j5/27 Introduction Cont.

Gliomas

Abnormal growth of cells.

Brain and spinal cord

tumors(Pereira et al. 2016).

Graded as Low Grade Gliomas(LGG)

and High Grade Gliomas(HGG). Gliomas (Pereira et al. 2016)

2019 Tumor Segmentation Using CNN

j6/27 Introduction Cont.

Manual segmentation(Forstmann, Keuken, and Alkemade

2015).

Semi-automatic or automatic methods(Forstmann, Keuken, and

Alkemade 2015).

Probabilistic atlases(Menze, Van Leemput, et al. 2010).

Tumor growth models(Menze, Van Leemput, et al. 2010).

From Voxel Distribution(Menze, Van Leemput, et al. 2010).

Using classifers like support vector machines / Random

Forests/ Spatially Adaptive RF(Bauer, Nolte, and Reyes 2011).

Deep learning(Bauer, Nolte, and Reyes 2011).

2019 Tumor Segmentation Using CNN

j7/27 Proposed methodology

Overview of proposed method (Pereira et al. 2016)

2019 Tumor Segmentation Using CNN

j8/27 Proposed methodology

A:P re processing

Remove bias felds

distortion

Apply intensity

normalization

Compute mean

intensity value and

Standard deviation

Normalize the patches Neural Bias feld distortion (Ny



ul,

Udupa, and Zhang 2000) Intensity normalization (Ny



ul, Udupa,

and Zhang 2000)

2019 Tumor Segmentation Using CNN

j9/27 Proposed methodology

B:C onvolutionalN euralN etwork

(Krizhevsky, Sutskever, and G. E. Hinton 2012) Initialization :

Done to make sure that the input values fall inside a desired

normal distribution.

Xavier initialization is used (Glorot and Yoshua Bengio 2010).

Activation Function :

Act as a gate between current layer and next layer.

Rectifer linear units are used(Krizhevsky, Sutskever, and

G. E. Hinton 2012).

Pooling :

Used to reduce the total computational load and represent

details in a more compact way.

Max pooling is used(LeCun, Bengio, and G. Hinton 2015).

2019 Tumor Segmentation Using CNN

j10/27 Proposed methodology

Regularization :

To generalize the model which makes less prone to

overftting(Krizhevsky, Sutskever, and G. E. Hinton 2012).

Remove nodes with some probability, Dropout(Srivastava et al.

2014) is used.

Data Augmentation :

Helps to artifcially expand the training size(Krizhevsky,

Sutskever, and G. E. Hinton 2012).

By rotating the original data by 90 degree.

Loss Function :

Indicates how well an algorithm models the data set.

Categorical cross-entropy is used.

2019 Tumor Segmentation Using CNN

j11/27 Proposed methodology

Loss Function :

For each data value specifed category is used.

It compare the predicted value with the true distributions, where

the solution is set to either 1 or 0.

H= X

j voxels X

kclasses c

j;k log

(c 0

j;k

)

(1)

where c’ is probabilistic predictions and c, the target.

Training :

Stochastic Gradient Descent, which actually replaces the actual

gradient and is used to minimize the lose function.

At region of low curvature Nesterov’s accelerated Momentum is

used to accelerate the algorithm.

2019 Tumor Segmentation Using CNN

j12/27 Proposed methodology

C:P ost P rocessing

Impose Volumetric Constrains-di erential and remove tissue clusters

and tumor cells Clusters miss-interpreted as tumor cells (Ny



ul, Udupa, and Zhang 2000)

2019 Tumor Segmentation Using CNN

j13/27 Experimental setup

A:Database

BRATS 2013 and 2015 databases are used for

validation(Menze, Jakab, et al. 2014).

4 MRI sequences are available for each patient .

BRATS 2013 have three data sets: Training, Leader-board and

Challenge.

BRATS 2015 have the Training set of 220 HGG and 54 LGG.

2019 Tumor Segmentation Using CNN

j14/27 Experimental setup

B:S etup

Tumor is approached as a multi-class classifcation problem.

For training the Neural network 450,000 HGG and 335,000 LGG

patches where extracted.

CNN is developed using Theano(Bastien et al. 2012) and

Lasagne(Dieleman et al. 2015).

2019 Tumor Segmentation Using CNN

j15/27 Experimental setup

C:E valuation

Considered 3 metrices: 1 Dice Similarity Coe cient (DSC):measures the overlap between

the manual and the automatic segmentation

DS C= 2

T P F P

+ 2T P +F N (2)2 Positive Predictive Value (PPV): measure of the amount of FP

and TP

P P V=T P T P

+F P (3)3 Sensitivity: evaluate the number of TP and FN detections

Sensitivity=T P T P

+F N (4)

where TP, FP and FN are the numbers of true positive, false

positive and false negative detections, respectively.

2019 Tumor Segmentation Using CNN

j16/27 Results and Discussions

Validation of Key Components. 1 Pre-processing

2 Data Augmentation

3 Activation Function

4 Deeper Architectures/Small Kernels

Patch Extraction Plane.

Global Validation.

2019 Tumor Segmentation Using CNN

j17/27 Results and Discussions

Validation of Key Components

Mean gain is calculated by subtracting the metrics of alternating

method and proposed method.

Leader-board data set. Diamonds indicate mean(Pereira et al. 2016)

.

2019 Tumor Segmentation Using CNN

j18/27 Results and Discussions

Validation of Key Components

Proposed pre-processing increased the detection of the complete

tumor by improving segmentation.

Challenge data set. Diamonds gives the mean (Pereira et al. 2016)

2019 Tumor Segmentation Using CNN

j19/27 Results and Discussions

Validation of Key Components

Over-segmentation of tumor gives these images. In HGG variant 2

classifed some non-enhanced tumor inside enhancing ring and in

LGG even in big Kernals couldn’t give total enhancment.

HGG Segmentation with cross-validation showing e ect of each component of the

proposed methods. Di erent colours represent tumor class. (Pereira et al. 2016)

.

2019 Tumor Segmentation Using CNN

j20/27 Results and Discussions

Validation of Key Components

LGG Segmentation with cross-validation showing e ect of each component of the

proposed methods. Di erent colours represent tumor class. (Pereira et al. 2016)

.

2019 Tumor Segmentation Using CNN

j21/27 Results and Discussions

Global Validation

Fig shows T1,T1c,T2,FLAIR and the segmentation. Each colour

gives a tumor class.

HGG and LGG Segmentation in the Leader-board data set (Pereira et al. 2016)

.

2019 Tumor Segmentation Using CNN

j22/27 Results and Discussions

Global Validation

A patient with 2 tumor were correctly detected and segmented from

Challenge data set.

Segmentation in Challenge data set. From left: T1,T1c,T2,FLAIR and segmentation.

(Pereira et al. 2016)

.

2019 Tumor Segmentation Using CNN

j23/27 Conclusions

Proposed method helps to reduce the computation time as well as

increase image enhancement comparing with other methods.

No limitations or disadvantages regarding method are discussed in

the paper.

2019 Tumor Segmentation Using CNN

j24/27 QUESTIONS?

2019 Tumor Segmentation Using CNN

j25/27 Literature I

Bastien, Fr

ed

eric et al. (2012). Theano: new features and speed

improvements”. In: arXiv preprint arXiv:1211.5590 .Bauer, Stefan, Lutz-P Nolte, and Mauricio Reyes (2011). Fully

automatic segmentation of brain tumor images using support

vector machine classifcation in combination with hierarchical

conditional random feld regularization”. In: international

conference on medical image computing and computer-assisted

intervention . Springer, pp. 354{361.Dieleman, S et al. (2015).

Lasagne: First release., doi:

10.5281/zenodo. 27878 .Forstmann, Birte U, Max C Keuken, and Anneke Alkemade (2015).

An introduction to human brain anatomy”. In: An Introduction

to Model-Based Cognitive Neuroscience . Springer, pp. 71{89.

2019 Tumor Segmentation Using CNN

j26/27 Literature II

Glorot, Xavier and Yoshua Bengio (2010). Understanding the

di culty of training deep feedforward neural networks”. In:

Proceedings of the thirteenth international conference on artifcial

intelligence and statistics , pp. 249{256.Krizhevsky, Alex, Ilya Sutskever, and Geo rey E Hinton (2012).

Imagenet classifcation with deep convolutional neural networks”.In: Advances in neural information processing systems ,

pp. 1097{1105. LeCun, Y, Y Bengio, and G Hinton (2015). Deep learning. nature

521″. In: Menze, Bjoern H, Andras Jakab, et al. (2014). The multimodal

brain tumor image segmentation benchmark (BRATS)”. In: IEEE

transactions on medical imaging 34.10, pp. 1993{2024.

2019 Tumor Segmentation Using CNN

j27/27 Literature III

Menze, Bjoern H, Koen Van Leemput, et al. (2010). A generative

model for brain tumor segmentation in multi-modal images”. In:

International Conference on Medical Image Computing and

Computer-Assisted Intervention . Springer, pp. 151{159.Ny



ul, L

aszl



o G, Jayaram K Udupa, and Xuan Zhang (2000). New

variants of a method of MRI scale standardization”. In: IEEE

transactions on medical imaging 19.2, pp. 143{150.Pereira, S

ergio et al. (2016). Brain tumor segmentation using

convolutional neural networks in MRI images”. In: IEEE

transactions on medical imaging 35.5, pp. 1240{1251.Srivastava, Nitish et al. (2014). Dropout: a simple way to prevent

neural networks from overftting”. In: The journal of machine

learning research 15.1, pp. 1929{1958.