A New Deep Convolutional Neural Network Model

for Automated Breast Cancer Detection

Xujuan Zhou Yuefeng Li Raj Gururajan
School of Management and Enterprise Science and Engineering Faculty School of Management and Enterprise
University of Southern Queensland Queensland University of Technology University of Southern Queensland
Springfield, Australia Brisbane, Australia Springfield, Australia
[email protected] [email protected] [email protected]

Ghazal Bargshady Xiaohui Tao Revathi Venkataraman

School of Management and Enterprise School of Sciences Department of Computer Science and
University of Southern Queensland University of Southern Queensland Engineering
Springfield, Australia Toowoomba, Australia SRM Institute of Science and Technology,
[email protected] [email protected] India
[email protected]
2020 7th International Conference on Behavioural and Social Computing (BESC) | 978-1-7281-8605-4/20/$31.00 ©2020 IEEE | DOI: 10.1109/BESC51023.2020.9348322

Prabal D Barua Srinivas Kondalsamy-Chennakesavan

School of Management and Enterprise Rural Clinical School
University of Southern Queensland University of Queensland
Springfield, Australia Toowoomba, Australia
[email protected] [email protected]

Abstract— breast cancer is reported as one of most common data. It has become a mainstream machine learning technique
malignancy amongst women in the world. Early detection of this with capacity in various nonlinear modelling tasks such as the
cancer is critical to clinical and epidemiologic for aiding in classification and feature extraction process from complex
informing subsequent treatments. This study investigates datasets. In recent years, many deep learning techniques
automated breast cancer prediction using deep learning including the convolutional neural networks (CNN) [6, 7],
techniques. A new 19-layer deep convolutional neural network and recurrent neural networks (RNN) [8], have been explored
(CNN) model for detecting the benign breast tumors from for classification tasks and applied in cancer detection
malignant cancers was proposed and implemented. The research field.
experiments on BreaKHis dataset was conducted and K-fold
Cross Validation technique are used for the model evaluation. In this study, a 19-layer deep convolutional neural
The proposed 19-layer deep CNN based classifiers compared network (CNN) model is proposed and employed to
with conventional machine learning classifier, namely Support automatically detect the breast cancer from mammography
Vector Machine (SVM) and a state-of-the-art deep learning images to give clinicians a powerful way of predicting patient
model, namely GoogLeNet in terms of Accuracy, Area under outcomes. The new model was evaluated on the real world
the Receiver Operating Characteristic (ROC) Curve (AUC), the dataset and the performance of the proposed model better
Classification Mean Absolute Error (MAE), Mean Squared than other machine learning and deep learning based models.
Error (MSE) metrics. The results demonstrate that the
proposed new model outperformed the other classifiers. The The rest of the paper is organized as follows: Section II
proposed model achieved an accuracy, AUC, MAE and MSE of briefly discusses related works. In section III, the proposed
84.5%, 85.7%, 0.082, and 0.043, respectively. 19-layer deep convolutional neural network (CNN) model is
introduced. Then the experimental results on BreaKHis
Keywords — breast cancer; deep convolutional network, dataset and discussions about the obtained outcomes are
machine learning, deep learning, computer vision presented in Section IV. The conclusion and some future
I. INTRODUCTION works are provided in Section V.
Cancer is one of the major causes of mortality worldwide. II. RELATED WORKS
Approximately 1 in 37 breast cancer patients will die as a In the last few years, Deep learning (DL) has become the
result of the disease and it has been cited as the second most classifier of choice in many machine learning tasks including
common cause of cancer-related death amongst females [1]. cancer prediction. Convolutional Neural Networks (CNN) is
In fact, an intelligent automated cancer prediction system can a subset of deep learning and it is a special kind of multi-layer
help for early detection and hence reduce cancer mortality. neural networks which is designed to recognize visual
Machine-learning (ML) techniques have been applied for patterns directly from pixel images with minimal pre-
cancer detection in the past decades. For example, Abdar et. processing. In recent year, CNN has attracted a lot of
al. [2] proposed and implemented a nested ensemble model attention and a variety of CNN models have been applied for
and achieved 98.07% accuracy. In [3], a confidence- cancer and cancer recurrence detection. In the following, we
weighted, voting and boosting ensemble techniques applied briefly introduce some cancer prediction research works
to combines boosting ANNs and two SVMs for BC diagnosis. using CNN techniques.
Deep learning [4, 5] is a machine learning technique Kumar et al. [9] proposed a prostate cancer recurrence
based on representation learning where the system approach base on deep learning technique. In their new
automatically learns and discovers the features needed for approach, two separate convolutional neural networks
classification from the processing of multiple layers of input (CNNs) were used. One is used to detect individual nuclei

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

Authorized licensed use limited to: University of New South Wales. Downloaded on May 18,2021 at 09:00:12 UTC from IEEE Xplore. Restrictions apply.
even in the crowded areas, and the other to classify them. parameters. In this study, the max-pooling operation is used.
Ciresan et al. [10] used deep max-pooling convolutional Max-pooling operation selects only the maximum value in
neural networks for mitosis detection for primary breast each feature map and consequently reducing the number of
cancer grading. Xu et al. [11] developed a Deep output neurons.
Convolutional Neural Networks (DCNN) based feature
learning method to automatically segment or classify The fully connected layer has full connection to all the
Epithelial (EP) and stromal (ST) regions from digitized tumor activations in the previous layer. A fully-connected layer
tissue microarrays (TMAs) for cancer detection. Litjens et al. includes neurons that each see and transform the entire input
[12] show the deep learning techniques could potentially volume to a one-dimensional vector. The dimension of its
reduce the workload for pathologists, while at the same time weight matrix is the dimension of flattened input multiple
to increase objective of diagnoses through two examples numbers of neurons Dflat×K and the output volume for an
which are prostate cancer and breast cancer. Samala et. al. input of size dimension is: N×K, where Dflat is dimension of
[13] developed an efficient deep convolutional neural network each image (usually 3 for color image), K is number of
model by reducing the number of computations to perform neurons and N is the numbers of samples.
breast cancer binary classification using all types of The activation function of a node defines the output of
Mammograms image data. Zia et.al. [14] proposed a multi- that node given an input or set of inputs. In this work,
deep model for lung nodule classification. The proposed Sigmoid activation function is used. Sigmoid activation
model consists of three phases including multi-scale dilated functions use a second non-linearity for large inputs. The
convolutional blocks (MsDc), dual deep convolutional neural layer 19 uses Sigmoid function to predict which class (benign
networks (DCNN A/B), and multi-task learning component and malignant) the input image belongs to. In this study, the
(MTLc). fully connected layer is a dense layer 1024 with sigmoid
activation for the binary classification of cancer images.
III. PROPOSED FRAMEWORK OF DEEP NEURAL NETWORK
In this study, 19-layer deep convolutional neural network IV. EXPERIMENTS AND RESULTS
framework was proposed and developed for detecting the To evaluate the proposed new deep learning model, the
benign breast tumors from malignant cancers. Fig. 1 shows experimental tests are conducted. The experimental details
the architecture of the proposed model. As shown in Fig. 1, including study dataset, experimental results, and the
the proposed model has two procedures. The first process is comparisons of the performance with other models are
data pre-processing, and second process is applying a fine- explained in the following subsections.
tuned 19-layer deep CNN to extract features and then classify
the breast cancer images into benign or malignant class. A. Dataset
The BreaKHis dataset [16], which is a publicly available from
A. Pre-Processing http://web.inf.ufpr.br/vri/breast-cancer-database, was used
In order to feed the images data into a new developed 19- for experimentation and evaluation of the newly proposed
layer deep CNN, the original images sizes have been changed model. The BreaKHis database has been built in
to 224×224×3. The original images contain black borders on collaboration with the P&D Laboratory — Pathological
both the left and right sides and text annotations in the upper Anatomy and Cytopathology, Parana, Brazil and it contains
left corner. To remove these undesired areas, the resulting microscopic biopsy images of benign and malignant breast
images are cropped and saved in three-channel RGB. tumors. BreaKHis consists of 7,909 clinically representative
microscopic images of breast tumor tissue images collected
B. 19-layer deep CNN netowrk
from 82 patients using different magnifying factors (40×,
Inspired by the state-of-the-art VGG16 deep 100×, 200×, and 400×) [16]. The details of the database can
convolutional neural network model that was introduced by be found in [16]. Fig. 2 and Fig. 3 show the samples of the
Simonyan and Zisserman in [15], in this study a new 19- layer benign and malignant classes in the BreaKHis dataset.
deep convolutional neural network model was proposed for
breast cancer detection. The proposed CNN architecture The dataset was split to benign and malignant sets and
includes 13 convolutional, 5 max-pooling, and 1 fully there are 2480 benign and 5429 malignant samples. The
connected layers and in total of 19 layers deep CNN. dataset is unbalanced. Therefore, a resampling technique by
copying and increasing the samples in the class with less
In general, the CNN architecture consists of three samples (i.e., benign class) to balance the number of samples
different types of layer: convolutional layer, pooling layer, for each class. Accordingly, there are 5,429 images for
and a fully connected layer. The convolutional layer performs malignant class and 4,960 images for benign class, and totally
the convolution on the input image data with the kernel using 10,389 images are applied after resampling (see Table I).
equation (1), where x is images, h is filter, and N is the
number of elements in x. The output vector is y. The Table I. The number of samples in dataset after and before
subscripts denote the nth element of the vector. resampling per class.
Benign Malignant Total images
= ∑ ℎ (1) Original 2,480 5,429 7,909
dataset
The pooling layer is used to deal with the overfitting
problem. The pooling operation reduces the dimension of Resampled 4,960 5,429 10,389
output neurons from the convolutional layer to reduce the dataset
computational intensity and decrease the number of

Authorized licensed use limited to: University of New South Wales. Downloaded on May 18,2021 at 09:00:12 UTC from IEEE Xplore. Restrictions apply.
 Fig. 1. The architecture of the proposed deep neural network model for breast cancer detection

B. Results The performances of the proposed prediction model was

In this study, the proposed 19-layer deep CNN based evaluated using the typical metrics for data classification
algorithm are built under an Intel core i7 @ 3.3 GHz and 16 including Accuracy, Area under the Receiver Operating
GB memory computer. Python software [17] was used for the Characteristic (ROC) Curve (AUC), the Classification Mean
model construction and prototyping, since it has freely Absolute Error (MAE), Mean Squared Error (MSE) metrics.
available libraries suits for deep learning. The parameters A conventional backpropagation (BP) [18] with a batch
were set as epoch =5 and batch size = 48. A 10-fold cross size of 48 is employed in this study to train the proposed
validation method was used for evaluate the performance of CNN. BP is a method to calculate the gradient of the loss
the model. Therefore, a total of 50 epochs of training were function with respect to the weights. BP passes error images
run in the experiment. An epoch refers to one iteration of the backwards through the network during training in order for
full training set. the weights to get updated to the network. A batch size is the
number of images used for each training update.
For the 10-fold cross-validation the sample image
datasets are randomly divided into ten equal parts. Nine out
of ten parts of breast images dataset are used to train the CNN
while the remaining one-tenth of the breast images dataset are
used to test the performance of the model. This strategy is
repeated ten times by shifting the test and training dataset.
The accuracy, AUC, MAE and MSE values reported in the
paper are the average values obtained from ten evaluations.
The results are indicated in Table II.
As can be observed from the Table II, the proposed model
shows good performance in terms of accuracy and AUC and
achieved 84.5% accuracy and 85.7% AUC. The proposed
model performed effectively in terms of MSE and MAE
measured for training and validation accuracy by low error
for MSE = 0.043, and MAE = 0.082.
Fig. 2. Samples of the Benign classes of the BreaKHis
dataset [16]. C. Comparison with the other models
Finally, we compare our proposed model with
conventional machine learning classifier, namely Support
Vector Machin (SVM) and a state-of-the-art deep learning
model, namely GoogLeNet.CNN model.
SVM [19] is a very popular classification algorithm. It has
been applied in a number of cancer prediction applications.
GoogLeNet, also known as Inception v1, was proposed by the
researchers at Google in 2014 [20]. The model is comprised
of a basic unit referred to as an "Inception cell". A series of
convolutions at different scales is performed and the results
are subsequently aggregated in Inception modules.
The SVM and GoogLeNet based classifier are built under
the same experimental setting with the proposed deep CNN
classifier. The comparison results are displayed in Table III.
As can be seen from Table III, the proposed method
significantly outperforms the other two baseline models in
Fig. 3. Samples of the Malignant classes of the BreaKHis terms of accuracy and AUC.
dataset [16].

Authorized licensed use limited to: University of New South Wales. Downloaded on May 18,2021 at 09:00:12 UTC from IEEE Xplore. Restrictions apply.
 V. CONCLUSION AND FUTURE WORKS [3] Abdar, M. and Makarenkov, V., 2019. CWV-BANN-SVM ensemble
learning classifier for an accurate diagnosis of breast cancer.
A 19-layer deep convolutional neural network (CNN) is Measurement, 146, pp.557-570.
developed to classify the benign and malignant class in this [4] LeCun, Y., Bengio, Y. and Hinton, G., 2015. Deep learning. nature,
study. A 19-layer CNN is proposed as it provides good 521(7553), pp.436-444.
convergence and the highest performance accuracy. All [5] Goodfellow, I., et al., Deep learning. Vol. 1. 2016: MIT press
parameters of the CNN structure are carefully fine-tuned in Cambridge.
order to obtain a model with optimal convergence rate. [6] LeCun, Y., Boser, B., Denker, J.S., Henderson, D., Howard, R.E.,
Hubbard, W. and Jackel, L.D., 1989. Backpropagation applied to
The proposed model was evaluated on the BreaKHis handwritten zip code recognition. Neural computation, 1(4), pp.541-
database after balancing the database. The data resampling 551.
applied in the pre-processing phase to improve results. The [7] LeCun, Y., Bottou, L., Bengio, Y. and Haffner, P., 1998. Gradient-
experimental results demonstrate that the proposed 19-layer based learning applied to document recognition. Proceedings of the
IEEE, 86(11), pp.2278-2324.
deep CNN model for breast cancer detection has a better
[8] Bengio, Y., Goodfellow, I. and Courville, A., 2017. Deep learning (Vol.
performance than a traditional classifier SVM and a deep 1). MIT press.
CNN GoogLeNet based classifier. [9] Kumar, N., Verma, R., Arora, A., Kumar, A., Gupta, S., Sethi, A. and
In the future, more study will be on how to advance the Gann, P.H., 2017, March. Convolutional neural networks for prostate
cancer recurrence prediction. In Medical Imaging 2017: Digital
proposed deep convolutional neural network model and test Pathology (Vol. 10140, p. 101400H). International Society for Optics
on different cancer image databases for real-time applications. and Photonics.
Another research direction in the future is to investigate more [10] Cireşan, D.C., Giusti, A., Gambardella, L.M. and Schmidhuber, J.,
efficient feature extraction method that can extract the features 2013, September. Mitosis detection in breast cancer histology images
that matter for the cancer image classification. with deep neural networks. In International Conference on Medical
Image Computing and Computer-assisted Intervention (pp. 411-418).
Springer, Berlin, Heidelberg.
Table II. The average performance of the proposed
[11] Xu, J., Luo, X., Wang, G., Gilmore, H. and Madabhushi, A., 2016. A
model for 10 runs. deep convolutional neural network for segmenting and classifying
epithelial and stromal regions in histopathological images.
MSE MAE Accuracy AUC
Neurocomputing, 191, pp.214-223.
[12] Litjens, G., Sánchez, C.I., Timofeeva, N., Hermsen, M., Nagtegaal, I.,
0.043 0.082 84.5% 85.7% Kovacs, I., Hulsbergen-Van De Kaa, C., Bult, P., Van Ginneken, B.
and Van Der Laak, J., 2016. Deep learning as a tool for increased
accuracy and efficiency of histopathological diagnosis. Scientific
reports, 6, p.26286.
[13] Samala, R.K., Chan, H.P., Hadjiiski, L.M., Helvie, M.A., Richter, C.
Table III. Comparison with baselines models in breast and Cha, K., 2018. Evolutionary pruning of transfer learned deep
cancer classification in terms of Accuracy and AUC. convolutional neural network for breast cancer diagnosis in digital
breast tomosynthesis. Physics in Medicine & Biology, 63(9), p.095005.
Models Accuracy AUC [14] Zia, M.B., Zhou, J.J, Zhou, X., Xiao, N., Wang, J., and Khan, A., 2020.
Classification of Malignant and Benign Lung Nodule and Prediction of
GoogleNet 54.0% 58.0% Image Label Class using Multi-Deep Model. International Journal of
SVM 78.0% 70.0%% Advanced Computer Science and Applications, 11(3), pp 35-41.
[15] Simonyan, K. and Zisserman, A., 2014. Very deep convolutional
The proposed model 84.5% 85.7%% networks for large-scale image recognition. arXiv preprint
arXiv:1409.1556.
[16] Spanhol, F.A., Oliveira, L.S., Petitjean, C. and Heutte, L., 2015. A
ACKNOWLEDGMENT dataset for breast cancer histopathological image classification. IEEE
Transactions on Biomedical Engineering, 63(7), pp.1455-1462.
This research work is partially supported by the [17] Sanner, M.F., 1999. Python: a programming language for software
Commonwealth Innovation Connections Grant, Australia integration and development. J Mol Graph Model, 17(1), pp.57-61.
(No. RC54960). [18] Hung, S.L. and Adeli, H., 1993. Parallel backpropagation learning
algorithms on Cray Y-MP8/864 supercomputer. Neurocomputing,
REFERENCES 5(6), pp.287-302.
[19] Vapnik, V., 2013. The nature of statistical learning theory. Springer
[1] Dora, L., Agrawal, S., Panda, R. and Abraham, A., 2017. Optimal science & business media.
breast cancer classification using Gauss–Newton representation based [20] Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D.,
algorithm. Expert Systems with Applications, 85, pp.134-145. Erhan, D., Vanhoucke, V. and Rabinovich, A., 2015. Going deeper
[2] Abdar, M., Zomorodi-Moghadam, M., Zhou, X., Gururajan, R., Tao, with convolutions. In Proceedings of the IEEE conference on computer
X., Barua, P.D. and Gururajan, R., 2018. A new nested ensemble vision and pattern recognition (pp. 1-9).
technique for automated diagnosis of breast cancer. Pattern
Recognition Letters.

Authorized licensed use limited to: University of New South Wales. Downloaded on May 18,2021 at 09:00:12 UTC from IEEE Xplore. Restrictions apply.