You are on page 1of 12

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/368522280

Breast Cancer Detection Using Deep Learning

Chapter · February 2023


DOI: 10.1007/978-981-19-7982-8_8

CITATION READS

1 773

3 authors:

Bhavin Harish Gami Khushi Chauhan


University of Technology Sydney Charotar University of Science and Technology
1 PUBLICATION 1 CITATION 1 PUBLICATION 1 CITATION

SEE PROFILE SEE PROFILE

Brijeshkumar Y. Panchal
Sardar Vallabhbhai Patel Institute of Technology, Vasad (GTU)
26 PUBLICATIONS 27 CITATIONS

SEE PROFILE

All content following this page was uploaded by Brijeshkumar Y. Panchal on 16 February 2023.

The user has requested enhancement of the downloaded file.


Breast Cancer Detection Using Deep
Learning

Bhavin Gami, Khushi Chauhan, and Brijeshkumar Y. Panchal

Abstract Breast cancer is a very common type of cancer found in women. Approx-
imately, 43,000 deaths are recorded per annum worldwide due to breast cancer. With
the advancement in medical technology, computer-aided detection and diagnosis
(CAD) system is being used widely for the early detection of cancerous cells. Rapid
development in deep learning has made the task of detecting cancerous cells accurate
and trivial. In this paper, researcher used convolutional neural network (CNN) for
classifying cancerous cells. CNN is a type of neural network which is extensively
used for image processing, classification, and segmentation. The proposed system
has achieved 82% accuracy by successfully classifying cancerous cells into benign
and malignant which are the two common types of cancer cells found.

Keywords Histopathology · Benign/malignant · Classification · Deep learning ·


Receptive field · Convolution neural network

1 Introduction

Breast cancers are very life-threatening which are often shown up in women and
rarely in men too. There are two types of breast cancer: benign and malignant. Benign
tumors are not adverse. They are harmless and do not penetrate nearby tissues causing
any damage to other organs. Malignant tumors consist of cancerous cells which have
adverse effects on the body as they penetrate nearby tissues and in any part of the body
too. Mammography is a technique for observing cancerous cells. To help radiologists
and doctors successfully detect cancer, computer-assisted detection and diagnosis

B. Gami · K. Chauhan
Department of Computer Science and Engineering, Devang Patel Institute of Advance
Technology and Research (DEPSTAR), Faculty of Technology and Engineering (FTE), Charotar
University of Science and Technology (CHARUSAT), Anand, India
B. Y. Panchal (B)
Computer Engineering Department, Sardar Vallabhbhai Patel Institute of Technology (SVIT)-
Vasad, Gujarat Technological University (GTU), Anand, Gujarat, India
e-mail: panchalbrijesh02@gmail.com

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 85
N. Marriwala et al. (eds.), Mobile Radio Communications and 5G Networks,
Lecture Notes in Networks and Systems 588,
https://doi.org/10.1007/978-981-19-7982-8_8
86 B. Gami et al.

(CAD) software has been established and in clinical use meanwhile the 1990s. To
boost the CAD system, deep learning is being used. For effective treatment of breast
cancer, early detection is necessary. Various imaging techniques are used to detect
cancer, including mammography, ultrasound, thermography, etc. Among the above,
all imaging techniques mammography images are effectively used and give the best
results of imaging for detecting breast cancers. Masses and microcalcifications (MCs)
are two important early signs of the disease. The masses shown in figure A can be
benign or malignant. There is a 30% possibility that the doctor cannot detect the
signs of benign or malignant. In 1996, the researchers proposed a computer-assisted
detection and diagnosis (CAD) system which can classify normal and abnormal mass
using convolution neural networks (CNNs).

2 Literature Review

Research work was published in 2019 by Li Shen, Laurie R. Margolies, Joseph H.


Rothstein, Eugene Fluder, Russell McBride, and Weiva Sieh in which they have
worked on breast cancer detection using VGG16 (16 layers) and ResNet (50 layers)
convolution neural networks and have done detailed comparison on their results.
From the dataset of 2478 mammography images, two datasets were made, S1 and
S10, where S1 had a set of patches centered on ROI, while the S10 consists of
the whole mammogram. For ResNet, the highest patch classification accuracy was
0.87(AUC = 0.87) for S10 and was 0.85 highest for VGG16 in S10. The tests have
shown that the dataset with no centered ROI and the pre-trained weights gives the
best results [1].
Kalyani Wadka, Prashant Pathak, and Nikhil Wagh proposed an ANN model in
2019 that classified breast cancer and its performance analysis were done with an
SVM classifier. They used a dataset of 5000 images which was pre-processed firstly.
They have used various machine learning (KNN, SVM) and deep learning (ANN,
CNN, Inception V3) algorithms and compared their confusion matrix’s results. The
accuracy of SVM and ANN was found to be higher (SVM-95%||ANN-99%) [2].
Shwetha K, Sindhu S, Spoorthi M, Chaithra D published a paper in 2019 in which
they associated two models of CNN, InceptionV3, and MobileNet. They used two
different datasets, the 1st dataset had 2 classes only, while the 2nd dataset consisted
of 7 different classes. The data was pre-processed and trained with different models.
The results concluded that Inception V3 (83%) gave better results than MobileNet
(58%) model [3].
Saad Awadh Alanazi, M. M. Kamruzzaman, Md Nazrul Islam Sarker, Madallah
Alruwaili, Yousef Alhwaiti, Nasser Alshammari, and Muhammad Hameed Siddiqi
carried out research in 2021 in which they have trained machine learning models and
deep learning models for breast cancer classification for and compared them. They
have made 3 CNN models, one with 2 convolution layers, other two with 3 and 5
layered depth, respectively. The CNN model with 2 convolution layers (59%) was
Breast Cancer Detection Using Deep Learning 87

less accurate than the SVM model (78%), but by adding up the convolution layer,
the model attained the highest accuracy of 87% [4].
Essam Rashed and M. Samir Abou El Seoud published a paper in 2019 through
which researchers can conclude the importance of image segmentation in classifica-
tion problems. Here, they have used well-known CNN O-Net architecture, i.e., two
U-Net architecture (comprises of convolution and deconvolution layer). The dataset
consisting of 6,671 mammograms images was used. The classification accuracy for
microcalcification and masses is 94.31% and 95.01% [5].
Naresh Khuriwal and Nidhi Mishra proposed a CNN model in 2018 for breast
cancer classification. The dataset consisted of 200 mammogram images which were
pre-processed using watershed segmentation and texture segmentation. 12 features
were extracted from the pre-processed image which was given as an input to the
CNN model. The accuracy of the model was observed to be 98% [6].
Monika Tiwari, Praditi Shah, Rashi Bharuka, Reena Lokar published a paper in
2020 where they have worked on breast cancer classification using various machine
learning (SVM, KNN, random forest) and deep learning models (ANN, CNN, RNN)
and compared their results. They have used 30 extracted features from the 597 images
as a dataset. The input features were pre-processed and passed to different models.
They concluded that the deep learning algorithm ANN showed higher accuracy, i.e.,
99.3% than any other deep learning or machine learning models [7]. F.A. Spanhol,
L.S. Oliveira, C. Petitjean, C. and L. Heutte proposed a model in 2016 where they
have used AlexNet CNN architecture for binary classification of cancerous cells in the
breast. Firstly, the model was trained with LeNet architecture, but it provided indef-
inite results. Then, they worked on AlexNet architecture where they have compared
two different patch extraction strategies at the convolution layer, i.e., sliding window
and random extraction. They concluded that AlexNet provides better results than the
class CNN, and the random extraction method of convolution comes out to be 89.4%
accurate [8].

3 CNN Architecture Theory and Dataset Used

Among the various datasets available for the histopathology of breast cancer in this
work, researchers have used the breast cancer histopathological image classification
(BreakHis) dataset collected by P&D Laboratory, Brazil. It consists of 9019 samples
of breast tumor tissues (mammography) of 82 patients with various magnifying ratios,
i.e., 40×, 100×, 200×, and 400× zoom and have RGB color code scheme. From this
9019, 2480 are benign type tumors and 5429 malignant type tumors. Our dataset used
only 1305 samples consisting of 407 benign samples and 898 malignant samples.
Convolution neural network is a type of artificial neural network (ANN) widely
used for image classification and processing which is specially designed to deal with
pixel data. CNN comprises various layers which are not present in the classic neural
network which gives the CNN architecture advantage of reducing the image in a
specific way that it does not lose its features and is made easier to process [9]. These
88 B. Gami et al.

layers consist of artificial neurons. Neurons are responsible for the feature extraction
in various layers available. When an image is specified as an participation to the
CNN model, pixel patches of images are provided to the neurons which then detects
features. The first layer detects the basic features of the image such as orientation and
size. The obtained output then goes to a further layer, and higher-level features are
extracted as the layer number increases. The figure shows different layers in CNN
architecture [10]. The convolution layer is responsible for convolution which is a
mathematical operation done between two signals which produce a new signal. In
image processing, the image is convoluted by the kernel/filter by placing a kernel
on each pixel of the image and getting a single-dimensional value for that particular
receptive field. By this, researchers get a reduced image with extracted features.
Talking about the pooling layer, it performs operations to minimize the dimensions
of the input/extracted feature array which in turn minimizes the computation cost
[11]. Max pooling and average pooling are the types of it. Similar to convolution, a
kernel is placed on the input, and a single value is obtained for that particular part of
the input. Based on the type of pooling, the value will be obtained. For max pooling,
the maximum value in the overlapped part of the kernel and input will be the single
value output for that part, and this happens for all the segments of input by placing
the kernel one by one in each part. For average pooling, the average of the values in
the overlapped part of input and kernel will be taken (Fig. 1).
The pooling layer also performs noise reduction in the inputs. The definite output
of classification will be provided by the last fully connected layer that provides the
probabilistic value in which neurons in the hidden layers are connected neurons.
This fully connected layer easily classifies inputs with non-linearity. The input that
is given to this fully linked layer is first flattened into a one-dimensional array by
the flatten layer just before the fully connected layer. Training is done in various
iterations over several epochs. The weight and bias are updated repeatedly for proper

Fig. 1 Two malignant tumor cells from the dataset


Breast Cancer Detection Using Deep Learning 89

Fig. 2 Two benign tumor cell from the dataset

tuning of the model. There are various activation functions available that are used to
produce output by taking weight and bias as input (Fig. 2).

4 Methodology

We have implemented CNN architecture implemented for image classification.


Figure 3 shows the amount of layers present in our proposed model. There are 15
layers in our model from which there are 3 convolution layers, 3 max pooling layers,
3 dense (hidden) layers, 3 dropout layers, and one flatten layer. Figure 4 shows the
flow of our proposed system. Firstly, researchers have resized and reshaped our data,
i.e., images. Then, inputs were normalized from 0–255 to 0–1. The images after
preprocessing are given as an input to the convolution layer which is the very first
layer of the architecture. As discussed in the above background theory, convolution
is used for feature extraction.
ΣΣ
G[m, n] = ( f ∗ h)[m, n] h[ j, k] f [m − j, n − k] (1)
j k

Equation 1 represents the convolution process in which f represents the input


image, h represents the kernel or filter, (m, n) is the size of the input image, and (j,
k) is the size of the filter.
Then, the output from this convolution passed to the max pool layer. All know
that the pooling layer performs down sampling, so the dimension of output from the
pooling layer will be given by this formula, W 2 × H 2 × D2 where W 2 = (W 1 −
F)/S + 1, H 2 = (H 1 − F)/S + 1, and D2 = D1 where F = filter size, S = stride,
W 1 , H 1 , and D1 are the dimension of input. Similarly, the features will be extracted
90 B. Gami et al.

Fig. 3 System flowchart

as the inputs get to the higher-level layer as per Fig. 3. As observed in Fig. 3, the
trainable parameters are increasing layer by layer which is nothing but the features
extracted from the images. The features extracted from the convolution and pooling
layer then are flattened into a 1D array and given as an input to the fully connected
layer 1. The fully connected layer 1 and 2 consist of 64 neurons and use rectified
linear unit (ReLU) activation function. ReLU activation function is widely used in
deep learning models. It ranges from 0 to infinity.
The fully connected layer 3 has only 2 neurons and uses the softmax activation
function which assigns probabilities (range [0–1]) to each class in a multi-class
problem like the classification problem. Figure 5 shows the inner structure of a
neuron. In our case, the inputs are the extracted features that are flattered in 1D-
array. Respective weights are multiplied with the input, and bias is added to it which
finally gives Σ(x i * W i + b). Then, this value is passed in an activation function
Breast Cancer Detection Using Deep Learning 91

Fig. 4 CNN architecture in our system

which in our case is ReLU which gives output Y = R(z) = {z for z > 0 and 0 for
z ≤ 0}. Finally, the fully connected layer floats over a value Y which shows the
probability of having any one type of cancer benign (0) or malignant (1).

Fig. 5 Artificial neuron structure


92 B. Gami et al.

Fig. 6 Predicted value of randomly picked image

5 Output Discussion

As seen in Fig. 6, researchers are providing a random malignant image as the input to
the model and it predicted it correctly as malignant with 87% accuracy. This model
is 83% accurate as shown in Fig. 4. Researchers have used 5 epochs and kept the
batch size 9 for training the model. They tuned the parameters like an epoch, batch
size, no. of layers in the architecture, the activation function, and the no. of neurons
while training the model. When researchers passed the validation data to the model,
it achieved 82% accuracy which was the maximum of all after tuning the model by
changing various parameters. Figure 7 displays the confusion matrix which displays
the representation of predicted values and true values.

6 Graph Discussion

Figure 8 shows the graph of training_loss versus validation_loss. Here, the loss repre-
sents how far is the projected value from the actual value. The first graph train_loss
versus validation loss represents how the loss function is declining with the rise no.
of epochs. In the training of the model with the zeroth epoch, the loss was greater
Breast Cancer Detection Using Deep Learning 93

Fig. 7 Confusion matrix

than 0.65 as shown in the graph. After validating the output with the true value, the
model back propagates and tunes the weights and bias in the first epoch for better
output.
As you can see in the graph, with the increase in an epoch, the loss is found to
be less than 0.60 in the first epoch. With the increase in no. of epoch, the model will
tune the weights and bias as per the requirement which will result in a decrease in the

Fig. 8 Loss function graph


94 B. Gami et al.

loss function. Similarly, the second graph shows training accuracy versus validation
accuracy when the no. of epoch increases. The graphs represent that with the increase
in the number of epochs, the loss decreases as we are predicting the values correctly.

7 Conclusion

Breast cancer detection using deep learning provides exceptional results while
performing precisely and accurately with proper data, fine-tuning the model, and
choosing the classification model. This artificial network is going to be way more
helpful shortly as it will be saving lives in numerous ways. Detecting breast cancer
is not just the single field where the neural network is serving. There are a lot more
medical diagnoses where it is used. The early detection of such life-threatening
diseases like breast cancer becomes an advancement in medical science as it will be
saving numbers of lives. This is just a single model which researchers have imple-
mented, there are various deep learning models providing almost 99% accuracy when
provided a mammographic image.

8 Future Work

The model has achieved 82% accuracy while the classification of breast cancer. It
uses approximately 35–40% of data available from the dataset. Shortly, researchers
will be using all of the datasets as contribution to the model. Additionally, the model
will be trained using various other CNN architectures for this full dataset such as
ResNet5, VGG16, and more for the comparison, and researchers are thinking of using
a classic machine learning classifier in place of the fully connected layer keeping the
other layers of CNN as it is for feature extraction.

References

1. Shen L, Margolies LR, Rothstein JH, Fluder E, McBride R, Sieh W (2019) Deep learning to
improve breast cancer detection on screening mammography. Sci Report 9(1):1–12. https://
doi.org/10.1038/s41598-019-48995-4
2. Wadkar K, Pathak P, Wagh N (2019) Breast cancer detection using ANN network and perfor-
mance analysis with SVM. Int J Comp Eng Technol 10(3):75–86, June 2019. https://doi.org/
10.34218/IJCET.10.3.2019.009
3. Shwetha K, Spoorthi M, Sindhu SS, Chaithra D (2019) Breast cancer detection using deep
learning technique. Int J Eng Res Technol (IJERT) 6(13):89–92. https://doi.org/10.1109/
EnCon.2019.8861256
4. Alanazi SA, Kamruzzaman MM, Sarker MNI, Alruwaili M, Alhwaiti Y, Alshammari N, Siddiqi
MH (2021) Boosting breast cancer detection using convolutional neural network. J Healthc Eng
2021:11. https://doi.org/10.1155/2021/5528622
Breast Cancer Detection Using Deep Learning 95

5. Rashed E, El Seoud MSA (2019) Deep learning approach for breast cancer diagnosis. In:
8th International conference on science and innovative engineering, pp 243–247, April 2019.
https://doi.org/10.1145/3328833.3328867
6. Khuriwal N, Mishra N (2018) Breast cancer detection from histopathological images using
deep learning. In: 2018 3rd International conference and workshops on recent advances and
innovations in engineering (ICRAIE), pp 1–4, November 2018. https://doi.org/10.1109/ICR
AIE.2018.8710426
7. Tiwari M, Shah P, Bharuka R, Lokare R (2020) Breast cancer prediction using deep learning
and machine learning techniques. SSRN:4–5
8. Spanhol FA, Oliveira LS, Petitjean C, Heutte L (2016) Breast cancer histopathological image
classification using convolutional neural networks. In: 2016 International joint conference on
neural networks (IJCNN), pp 2560–2567, July 2016. https://doi.org/10.1109/IJCNN.2016.772
7519
9. Shah R, Shastri J, Bohara MH, Panchal BY, Goel P (2022) Detection of different types of blood
cells: a comparative analysis. IEEE Int Conf Distrib Comp Electr Circ Electron (ICDCECE)
2022:1–5. https://doi.org/10.1109/ICDCECE53908.2022.9793132
10. Jayswal D, Panchal BY, Patel B, Acharya N, Nayak R, Goel P (2023) Study and develop
a convolutional neural network for MNIST handwritten digit classification. In: Singh PK,
Wierzchoń ST, Tanwar S, Rodrigues JJPC, Ganzha M (eds) Proceedings of third international
conference on computing, communications, and cyber-security. Lecture notes in networks and
systems, vol 421. Springer, Singapore. https://doi.org/10.1007/978-981-19-1142-2_32
11. Joshi DD, Joshi HH, Panchal BY, Goel P, Ganatra A (2022) A Parkinson disease classification
using stacking ensemble machine learning methodology. In: 2022 2nd International conference
on advance computing and innovative technologies in engineering (ICACITE), pp 1335–1341.
https://doi.org/10.1109/ICACITE53722.2022.9823509

View publication stats

You might also like