See discussions, stats, and author profiles for this publication at: [Link]
net/publication/349174886
Cancer Detection Using Convolutional Neural Network
Chapter · February 2021
DOI: 10.1007/978-3-030-67187-7_30
CITATIONS READS
10 3,806
3 authors, including:
Maheep Singh Krishan Kumar
National Institute of Technology (NIT) Uttarakhand National Institute of Technology (NIT) Kurukshetra
11 PUBLICATIONS 35 CITATIONS 81 PUBLICATIONS 1,377 CITATIONS
SEE PROFILE SEE PROFILE
Some of the authors of this publication are also working on these related projects:
LEXER: LEXicon based Emotion analyzeR View project
Affective computing View project
All content following this page was uploaded by Krishan Kumar on 09 June 2021.
The user has requested enhancement of the downloaded file.
Cancer Detection Using Convolutional Neural
Network
Abstract. Cancer is a dreadful heterogeneous disease that refers to ab-
normal growth of cell tissue. Millions of people die every year due to
unnoticed recognition or late detection of the disease. Hence, detecting
presence of cancer cells correctly becomes important. Putting into at-
tention the number of people affected by cancer, we need to find better
ways of diagnosis and treatment. Neural networks are a burning research
area in medical science, especially in the areas of radiology, cardiology,
oncology, urology and etc. A variety of these techniques as Decision Trees
(DTs), Artificial Neural Network (ANN) and Support Vector Machines
(SVM) have been widely applied in cancer detection. This paper presents
a method of classification of cancer cells into Benign and Malignant using
deep learning Convoluional Neural Network.
1 Introduction
Cancer is one of the major reason to cause death in the world [1]. American Can-
cer Society performed a survey that showed that around 600,920 people would
die from cancers in USA in 2017 [2]. Thus, fighting against cancers is one of
the important and a big challenge faced by both clinic doctors and research sci-
entists. According to a study for Breast Cancer Care, 42% of National Health
Service (NHS) trusts say that they do not have staffs to assign individual with
limited cancer specialist nurse [3]. It is a major reason that causes low survival
rate of cancer all over the world. Due to lack of specialist doctor or nurse, the
cancer remains undetected for a longer duration thus reducing chances of opti-
mal treatment and recovery. Early detection plays a key role in cancer diagnosis
and can lead to long survival rate. Manual interpretation of enormous medical
images is a tedious and time consuming task. Therefore, cancer detection tech-
niques using neural networks are being introduced that would result in early
diagnosis and treatment. Various feature extraction methods have been inves-
tigated for different cancer types. In recent years, advanced analysis of medical
imaging using machine learning, deep learning, including Convolutional Neural
Networks (CNNs), has been explored [4]. These approaches offer great promises
for future applications for both diagnostic and predictive purposes.
In this paper, we have used deep learning technology as the main algorithm
so as to recognize cancer cells. A classification occurs mainly when an object
needs to be assigned to a predefined group or class based on a number of ob-
served attributed related to that object. Mainly in this deep learning technology,
Convolutional Neural Network has been used for classifying the images. CNN is
basically a kind of multilayer perceptron that uses small sub-regions also called
2
as receptive fields. These fields are then tied to cover the entire image and then
produce the feature maps to share same weights and biases. Hence this archi-
tecture reduces the number of parameters which are needed for a neural netwok
learning process and thus increases the learning efficiency. Recognition of cancer
cells is usually solved in two steps- preprocessing the dataset images and sub-
sequent classification. For classification purpose, we have used ConvNet model
that was trained and tested on the Invasive Ductal Carcinoma (IDC) dataset [5].
For example-Figure 1. shows a cancer cell image.
The rest of the paper is organized as follows: Section 2 that provides informa-
tion about the related work done in this specific area. Section 3 gives description
about the architecture used to implement classification along with the proposed
implementation, while the Experimental Results carried out on Invasive Ductal
Carcinoma (IDC) Dataset are presented in Section 4. Finally the conclusions
and future work are drawn in Section 5.
Fig. 1: Cancer cell
2 Literature Survey
There are many researches in the literature dealing with Cancer detection and
recognition problem. Albayrak et al. developed a deep learning based algorithm
that used CNN model to extract features which were then used to train a Support
Vector Machine (SVM), to detect mitosis in breast histopathological images [6].
Spanhol et al. used AlexNet to classify benign and malignant tumors from the
histopathological images [7]. Pomponiu et al. proposed a skin classification al-
gorithm by applying a pre-trained CNN, AlexNet, to generate high-level feature
representation of the skin samples which where then used to train a k nearest
neighbour (KNN) classifier [8]. A deep cascade network was proposed by Chen
et al. for mitosis detection in breast histology slides [9]. They initially trained
a FCN model to extract mitosis candidates and then tuned a CaffeNet model.
Xu et al. proposed a Stacked Sparse Auto-Encoder (SSAE) algorithm to clas-
sify nuclei in cancer images [10]. Wichakam et al. proposed a system combining
Cancer Detection Using Convolutional Neural Network 3
SVM and deep CNN for mass detection on digital mammograms [11]. Along with
using pre-trained models, some CNN models have been developed for classifica-
tion. Demyanov et al. proposed using deep CNNs to detect two types of patterns
(typical network and regular globules) [12]. The CNN was trained by stochas-
tic gradient descent algorithm. Sabouri et al. presented a CNNs based border
detection system for skin lesions recognition. Swiderski et al. presented a way
to overcome the overfitting of CNN models when the data is limited. [13] They
used non-negative matrix factorization (NMF) and statistical self-similarity to
enrich the training data.
Apart from Convolutional Neural Netwok, various other deep learning tech-
niques have been used to implement the same on mammographic images. Dhun-
gel et al. proposed another algorithm for mass detection in mammograms. In the
proposed method, a cascade of random forest classifiers and deep learning were
used. Kim et al. proposed a 3-D multiview deep ConvoNet model to learn latent
bilateral feature representation of digital breast tomosynthessis (DBT) volume.
Two CNNs were used to extract higher level features from two VOIs seper-
ately. Masood et al. proposed a semi-supervised learning model for melanoma
detection. In the proposed system [14], a deep network and two support vector
machines (SVM) with radial basis function (RBF) kernel and polynomial ker-
nel were trained on three different datasets generated from both labeled and
unlabeled data. Majtner et al. proposed a classification system combining deep
features and hand-crafted features for melanoma recognition. Sabbaghi et al.
proposed a deep neural network that learned high-level image representation,
and mapped them into bag-of-features (BoF) space to enhance the classification
accuracy. Yu et al. proposed deep residual networks for recognizing melanoma
in dermascopy images using a Fully Convolutional Residual Network (FCRN).
3 Architecture
The architecture used in the present work departs from traditional ConvoNets
by the type and number of convolve and pooling layers used.
The architecture uses the ConvoNet architecture layers that are convolution
layer, dropout, fully connected layer for definite epochs to train the model.
3.1 Convolutional Neural Network
A neural network is a mathematical model based on neural units-artificial neurons-
similar to biological neural networks. Typically, neurons are organized in layers
and the connection are set up between neurons only from adjacent layers. The
input feature vector is put into first layer and, moving from layer to layer, is
transformed to a high level features vector. The output layer neurons finally
leads to classification and there number is equal to number of classifying classes.
Each convolutional layer consists of a set of trainable filters and computes dot
production between these filters and layer input to obtain an activation map.
4
Filters are also called kernels and allow detecting same features in different loca-
tions. The activation function used is RELU function and Sigmoid function [15].
Table 1: Layered Structure of proposed implementation
Layer Description
Input 50 × 50 × 3
Convolution 32 3 × 3, Valid padding, 1×1 stride
Relu
MaxPooling 2×2
Convolution 64 3 × 3, Valid padding, 1×1 stride
Relu
MaxPooling 2 × 2 , stride=2
Convolution 128 3 × 3, 1×1 stride
Relu
Convolution 256 3 × 3, 1×1 stride
Flatten
Dropout
Relu
Fully Connected 128
Relu
Dropout
Fully Connected 128
Relu
Fully Connected 2
Sigmoid function
3.2 Proposed Implementation
– Initially, all the dataset images are resized. This is done as few of the images
in dataset are not of size 50 × 50 × 3.
– Images are normalized by dividing it by 255 to ensure that all the values are
between 0 and 1. This is done to increase the model’s training efficiency and
to train the model faster. It also prevents us from falling into the exploding
gradients problem.
– Data imbalancing is handled by randomly undersampling the majority class.
The samples of the majority class are removed to make them equal to mi-
nority class.
– Data Augmentation is done to generalize the model by reducing network’s
capacity to overfit the training data. Rotation, horizontal and vertical flip-
ping techniques are performed to ensure the same. These steps are necessary
to ensure that all classes have sufficient number of images so that it is easy
for the network to predict all kinds of classes and not just a few of them on
Cancer Detection Using Convolutional Neural Network 5
which the network is trained.
Training process: We have used deep learning library Keras. Training and
testing was done using the Invasive Ductal Carcinoma (IDC) dataset. The devel-
oped method can classify the cells into benign and malignant. Table 1 describes
the network architecture. The architecture includes several convolutional layers,
Pooling, Fully connected, Dropout layers as shown in Table 1.
Convolution Layer: It comprises of a set of independent filters such that
each filter is independently convolved with the image.
m X
X m
hi,j = wk,l .xi+k−1,j+l−1 (1)
k=1 l=1
where, m-Kernel width and height
h-Convolution Output
x-Input
w-Convolution Kernel
Pooling Layer: It progressively reduces the size of representation to reduce
the amount of computation and parameters in the network. The approach used
here is MaxPooling [1].
hi,j = max{xi+k−1,j+l−1 ∀1 ≤ k ≤ m and1 ≤ l ≤ m} (2)
Relu and Sigmoid: These are the non linearity added to the network.
(
x, if x > 0
RELU = (3)
0, if x ≤ 0
1
Sigmoid = (4)
(1 + e−(ax) )
Dropout: To reduce the chances of overfitting, some of the neurons are dropped
[16].
Fully Connected Layer: This layer have full connections to all the activations
in the previous layer and their activation is computed with a matrix multiplica-
tion followed by a bias offset. This layer takes an input volume from the preceding
layers and outputs a n dimension vector where n is the number of classses.
Flatten: This layer is used to convert 3D feature maps to 1d feature vectors.
The model was trained in 80 epochs, using a batchsize of 256. The opti-
mizer used is Adam Optimizer to update weights and minimize loss, with initial
learning rate as 0.00001.
6
4 Experimental Evaluation and Results
This section present the experimental results carried out using ConvNets for
detecting cancer cell. We first discuss the dataset used for the experiment and
then the results. We have used Python language for the implementation of this
architecture.
4.1 Invasive Ductal Carcinoma Dataset Visualization
The Invasive Ductal Carcinoma Dataset is comprised of two different cancer cell
types sepcified as Benign and Malignant. Each image can be represented as a 50
× 50 × 3 array of RGB pixels in the range of 0 to 255. The details of dataset
are shown in Table 2. and Figure 2. shows the classes of the dataset.
Table 2: Dataset Summary
Class 0 44478
Class 1 15522
Shape of images 50 × 50 × 3
Number of unique classes 2
Fig. 2: Classes of Invasive Ductal Carcinoma dataset
4.2 Testing Results
Data Resizing and Data Augmentation was done on the dataset images to ensure
proper training efficiency. The training of the proposed model took 7 hours on
an intel CORE i3 using google colab platform. The model implementation gave
the following quantitative measures as shown in Table 3 and Figure 3 gives idea
Cancer Detection Using Convolutional Neural Network 7
about training and testing set loss:
Table 3: Experimental result
Batch size Epochs Test Accuracy
256 80 79.10%
Fig. 3: Training and Testing set loss
The confusion matrix is a binary classification matrix having four quadrants,
True Positive (Predicted malignant and actually malignant), False Positive (Pre-
dicted malignant but actually benign), False Negative (Predicted benign but
actually malignant), True Negative (Predicted benign and actually benign) as
shown in Table 4. Model’s performance is evaluated using confusion matrix as
shown in Table 5.
Table 4: Resultant Confusion Matrix
1761 551
416 1896
8
Table 5: Confusion Matrix representation in four quadrants
Predicted malignant and Predicted malignant and
actually malignant actually benign
Predicted benign Predicted benign
but actually malignant and actually benign
5 Conclusion
This paper presented an architecture that represents an effective way of detecting
cancer cells whether they are benign or malignant. The experimental results
achieved after testing on Invasive Ductal Carcinoma Dataset, conclude that the
model is invariant to scale and viewing angle. The quantitative results validated
that the model can be used to detect the cancerous cells on an average scale.
Apart from the promising results presented in the paper, there are different
aspects of future research. As a future work, the model can be enhanced by
using other non linear activation functions (such as ELU function) as well as
parallel layers that would result in an increase in accuracy.
References
1. Chen, W., Zheng, R., Baade, P.D., Zhang, S., Zeng, H., Bray, F., Jemal, A., Yu,
X.Q. and He, J., 2016. Cancer statistics in China, 2015. CA: a cancer journal for
clinicians, 66(2), pp.115-132.
2. Cancer facts & figures 2017, American Cancer Society, 2017, (2017)
3. Stevens, C., Vrinten, C., Smith, S.G., Waller, J. and Beeken, R.J., 2019. Accept-
ability of receiving lifestyle advice at cervical, breast and bowel cancer screening.
Preventive medicine, 120, pp.19-25.
4. Lambin, P., Rios-Velazquez, E., Leijenaar, R., Carvalho, S., Van Stiphout, R.G.,
Granton, P., Zegers, C.M., Gillies, R., Boellard, R., Dekker, A. and Aerts, H.J.,
2012. Radiomics: extracting more information from medical images using advanced
feature analysis. European journal of cancer, 48(4), pp.441-446.
5. Cancer Dataset: [Link]
histopathology-images
6. A. Albayrak, G. Bilgin, Mitosis detection using convolutional neural network based
features, in: Proceedings of IEEE Seventeenth International Symposium on Com-
putational Intelligence and Informatics (CINTI), 2016, pp.000335000340.
7. Krizhevsky, A., Sutskever, I. and Hinton, G.E., 2012. Imagenet classification with
deep convolutional neural networks. In Advances in neural information processing
systems pp. 1097-1105.
8. Shiraishi, J., Katsuragawa, S., Ikezoe, J., Matsumoto, T., Kobayashi, T., Komatsu,
K.I., Matsui, M., Fujita, H., Kodera, Y. and Doi, K., 2000. Development of a digital
image database for chest radiographs with and without a lung nodule: receiver
operating characteristic analysis of radiologists’ detection of pulmonary nodules.
American Journal of Roentgenology, 174(1), pp.71-74.
Cancer Detection Using Convolutional Neural Network 9
9. Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R., Guadar-
rama, S. and Darrell, T., 2014, November. Caffe: Convolutional architecture for
fast feature embedding. In Proceedings of the 22nd ACM international conference
on Multimedia pp. 675-678. ACM.
10. J. Xu, L. Xiang, R. Hang and J. Wu, Stacked Sparse Autoencoder (SSAE) based
framework for nuclei patch classification on breast cancer histopathology, 2014
IEEE 11th International Symposium on Biomedical Imaging (ISBI), Beijing, 2014,
pp. 999-1002.
11. I. Wichakam and P. Vateekul, Combining deep convolutional networks and SVMs
for mass detection on digital mammograms, 2016 8th International Conference on
Knowledge and Smart Technology (KST), Chiangmai, 2016, pp. 239-244.
12. S. Demyanov, R. Chakravorty, M. Abedini, A. Halpern and R. Garnavi, Classifica-
tion of dermoscopy patterns using deep convolutional neural networks, 2016 IEEE
13th International Symposium on Biomedical Imaging (ISBI), Prague, 2016, pp.
364-368.
13. D. H. Kim, S. T. Kim and Y. M. Ro, Latent feature representation with 3-D multi-
view deep convolutional neural network for bilateral analysis in digital breast to-
mosynthesis, 2016 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), Shanghai, 2016, pp. 927-931.
14. Mahbod, A., Schaefer, G., Wang, C., Ecker, R. and Ellinge, I., 2019, April. Skin
lesion classification using hybrid deep neural networks. In ICASSP 2019-2019 IEEE
International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp.
1229-1233. IEEE.
15. Hu, Z., Tang, J., Wang, Z., Zhang, K., Zhang, L. and Sun, Q., 2018. Deep learning
for image-based cancer detection and diagnosis A survey. Pattern Recognition, 83,
pp.134-149.
View publication stats