Page 66 - Big Data Analytics for Intelligent Healthcare Management

CHAPTER

TRANSFER LEARNING AND

SUPERVISED CLASSIFIER BASED 4

PREDICTION MODEL FOR

BREAST CANCER

Md. Nuruddin Qaisar Bhuiyan, Md. Shamsujjoha, Shamim H. Ripon, Farhin Haque Proma, Fuad Khan
Department of Computer Science and Engineering, East West University, Dhaka, Bangladesh

4.1 INTRODUCTION
There are several methods available for diagnosis of breast cancer such as breast exam, mammography,
breast ultrasound, and biopsy [1, 2]. Of these, biopsy is the only way to detect the presence of breast
cancer. The most common biopsy techniques are core needle biopsy, fine needle biopsy, and surgical
open biopsy. In the biopsy procedure, breast tissue samples are collected and examined under the
microscope by pathologists. The whole procedure is based on visual inspection by pathologists.
This is a time consuming, costly task and it requires attention of the pathologists examining the tissue.
Histopathology image analysis also depends on the experience of the pathologist, which is costly and
requires a huge amount of time [3].
As a result, there is a pressing need for an automated system that can differentiate between the can-
cerous tissue (malignant tissue) and the noncancerous tissue (benign tissue) and help the pathologists to
make the diagnosis process into an easy and time efficient task, and consequently the pathologist can
focus on more difficult cases.
A significant amount of research work has already been undertaken to build a computer aided
system to automate the classification of benign and malignant tumor tissue images using an image data-
set consisting of different sizes of images. To build an automated system, we used one of the largest
datasets of breast cancer images called BreaKHis containing 7909 images of benign and malignant
tissue at four different magnification factors. Convolution neural networks (ConvNets) are a state-
of-the-art technique for image classification. There are many available convolution networks released
by renowned organizations and institutions. Some architecture examples are ResNet-50, Inception V3,
Inception ResNet V2, Xception etc. These ConvNets are very deep and are trained on millions of im-
ages. Training these ConvNets from scratch requires a significant amount of time and can cause over-
fitting since the number of training images we have is not big enough. For this reason, in this paper, four
pretrained ConvNets were used as fixed feature extractors to extract the feature from the benign and
malignant tissue images. To reduce the dimension of extracted features from different ConvNet

59
Big Data Analytics for Intelligent Healthcare Management. https://doi.org/10.1016/B978-0-12-818146-1.00004-0
# 2019 Elsevier Inc. All rights reserved.