Scribd
Upload
4 views4 pages

05 Abstract

Uploaded by

gdheepak1979
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
4 views4 pages

05 Abstract

Uploaded by

gdheepak1979
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

ABSTRACT

Brain is a very complex and controlling central nervous


system organ of the human body. Abnormalities in brain may
lead to disturb the processing and coordinating activities of other
organs of the human body. Brain diseases can be identified
using different diagnosis methods. Brain tumors have been
identified in brain tissues as uncontrolled unusual cell growth.
Brain tumors are one of the most dangerous diseases, which
lead to reduce the life span of the patient or even lead to death.
Brain Tumor's Preliminary Diagnosis is a vital approach in
preparing a successful treatment plan. Throughout the
treatment of such essential diseases and in their further healing
process, medical imaging plays a significant role. Diagnosis can
be done through the medical brain imaging, which provides
structural and functional information of the brain. The
structural information provides the details about intracranial
diseases and nervous system. The functional information
provides the details of metabolic diseases and lesions on a finer
scale.

Diverse medical imaging modalities are available to acquire


a brain image. But MR (Magnetic Resonance) Imaging is more
popular, compared among all other medical imaging modalities
due to its superior quality of soft tissues and no ionizing
radiation. One more advantage of MR Brain imaging provides
different sub modalities like T1, T1c, T2, PD and FLAIR Images.
All these modalities used to identify the particular abnormality
in the brain. The radiologists investigate these medical images to
identify and classify the type of the tumor. This process will take
more time due to the hectic work schedule and non-availability
of Radiologists. The classification of brain tumors requires good

i
domain knowledge and there is an opinion mismatch between
the radiologists lead to a confused state in treatment planning.
The information collected by medical imaging machines includes
low-level information which human visual system cannot
perceive. As a remedy, classification of brain tumors by
computer-aided techniques decreases the diagnosis time and
provides an accurate tumor prediction and can serve as a second
opinion.

Machine learning algorithms have significantly contributed


to automating these tasks. These traditional machine-learning
algorithms are trained with handcrafted features extracted from
the MR Images. The feature selection depends on the domain
knowledge of the experts and it will change person to person.
Deep learning models have become more popular in the image
classification in recent years. CNNs have proven to be a master
at extracting the vast number of non-handcrafted features to
improve the accuracy of classification models. The deep learning
models will take an image as input instead of selected features.
In this work, a deep transfer-learning model adopted for the
classification of the brain tumors on two publicly available
datasets. The former one is BTDS-2 Dataset extracted from
BRATS 2018 database consists Benign and Malignant tumor
classes and other one is CE-MRI Dataset consists glioma,
meningioma and pituitary tumor classes. Pre-processing is
performed to improve the quality of the image for further
processing. DnCNN (Denoising CNN) and median filters are used
to remove the noise in MR brain images. Transfer learning model
is to transfer the knowledge of source image classification to
target image classification. Various pre-trained CNNs such as
AlexNet, VGG16, VGG19, ResNet50, ResNet101 and GoogLeNet

ii
are used for extraction of features, and newly specified fully
connected layers and classification softmax layer are widely
used. Accuracy, Sensitivity, Specificity, Precision and F1-Score
were determined as classification metrics. The VGG19 based
transfer-learning model achieved with outstanding accuracy of
92.37% for BDTS-2 Dataset and 95.43% for CE-MRI Dataset.

The features extracted from pre-trained CNN are fed to


softmax and classification layers in transfer learning model
previously. In this work, a hybrid network of pre-trained CNN
and a supervised classifier is trained to classify brain tumors.
The pre-trained CNN features are formed into a feature vector
corresponding to each MR brain image separately for both
training data set and testing data set. In examining the
performance of classification system, these feature vectors are
fed to train various machine learning classification algorithms
such as, Support Vector Machine (SVM), K-Nearest Neighbor (K-
NN) and Decision Tree classifiers. The results obtained are very
promising with CNN feature extraction. The proposed model is a
hybrid combination of CNN and KNN classifier with impressive
accuracy results of 96.1% for BDTS-2 Dataset and 96.74% for
CE-MRI Dataset.

In the next investigation stage to increase the correct


prediction rate, knowledge transfer with block wise fine-tuning of
pre-trained CNN networks can be implemented with separate
supervised classifier on small medical datasets for diagnosis. In
this research work, transfer learning and block wise fine-tuning
is used to classify datasets (BTDS-2 and CE-MRI) with separate
KNN classifier on pre-trained VGG deep neural network. The
method proposed is known as BT-VGGNet. The labeled BTDS-2

iii
and CE-MRI Datasets are exploited to block wise fine-tuning for
feature extraction through all hidden layers in the VGG deep
neural network and KNN for classification. Experimental results
have shown that the proposed model’s accuracy is outperformed
compared to the state-of-the-art classification methods with
97.28% for BDTS-2 Dataset and 98.69% for CE-MRI Dataset.

In the next analysis phase to increase the true prediction


rate for the MR brain tumor image classification, an automatic
computer-aided diagnostic system with the deep convolution
neural network trained from scratch is proposed can perform
well. However, with small datasets, over-fitting is one of the
major problems. To train a deep learning model from scratch it
requires large datasets. Generally, it is difficult to obtain a large
number of labelled samples in medical image classification
tasks. In this work, the dataset comprises 2D slices extracted
from BRATS 2015, BRATS 2017 and BRATS 2018 volumetric
data to feed proposed DL (Deep Learning) model training. The
proposed Tumor network (T-Net) is built using 16 layers from the
image input layer that holds the pre-processed images passing
through the layers of convolution and their activation functions.
The obtained results show a remarkable performance with a
classification accuracy of 97.96% for BDTS-2 Dataset and
99.34% for CE-MRI Dataset. It also improved the quality of
classification by increasing the performance metrics like
precision, sensitivity, specificity and F1-score.

iv

You might also like