2020 IEEE 20th International Conference on Communication Technology

Metastatic Cancer Image Binary Classification Based on Resnet Model

Mingrui Wang Xuhui Gong

College of Science University of Liverpool
Wuhan University of Technology Liverpool, England
Wuhan, China e-mail: [email protected]
e-mail: im [email protected]

Abstract—With the rapid development of computers technol- learning methods is labor intensive because of manual features
ogy, computers have inherent advantages over human in terms extraction [1], [2]. This methods has relatively low accuracy
of sample storage, data processing, and calculation speed. Com- and efficiency. However, deep learning methods can produce
puter assisted diagnosis achieves automatical disease detection
and classification which is economical. In recent years, using low-dimensional representations and capture specific features
deep learning methods to solve medical image classification effectively from image data types [3]–[6]. In particular, Con-
has received much attention from researchers due to great volutional Neural Networks (CNN) are widely used in medical
performance. In this article, we propose a new metastatic cancer image classification tasks [4].
image classification approach based on Resent model. We aim In this article, we aim to solve the small patch-level image
to turn the digital pathology scan detection task into a binary
classification task of cancer image. The ResNet model can effec- of metastatic cancer task which is provided by the Kaggle
tively alleviate the gradient explosion and gradient disappearance competition platform. We propose a new cancer image classi-
problems and train a deeper network by using a skip connection fication method based on Resent model which is an application
with an identity transformation. The experiments were conducted innovation. The ResNet (Residual Neural Network) [7] model
on the PCam benchmark dataset which is provided by the Kaggle ues a skip connection with an identity transformation which
competition platform. The experimental results indicate that our
method outperforms the other convolutional neural networks is proposed by He et al. This way could alleviate the problem
methods such as Vgg16 and Vgg19. Our proposed method has of gradient disappearance and gradient explosion to ensure
the highest scores on the Auc Roc and the Accuracy measures. good performance and train a deeper network. To evaluate
Experimental results show that the use of deep learning method our method, we conduct experiments on the PCam benchmark
based on Resent can significantly improve the performance of dataset by Auc Roc and Accuracy metrics. Our PatchCame-
metastatic cancer diagnosis.
lyon (Pcam) dataset is small metastatic cancer image patches
Keywords—Resent Model; Cancer Image Classification; Con- which are got from large pathology scan images. Fig. 1 is
volutional Neural Networks the histopathological scans of lymph node section from our
dataset. The first row is positive samples and the tag of positive
I. INTRODUCTION sample is 1 which represents a cancer. The second row is
Through medical knowledge and years of medical learning, negative samples and the tag of negative sample is 0 which
human doctors can use the medical images obtained by MRI represents not a cancer. Experimental results demonstrate our
or CT to find relevant disease characteristics, and make cor- ResNet50 method achieves best performance than the VGG16
responding medical diagnosis and treatment plans. However, and VGG19 [8]. Specifically, the Auc Roc score and the
this way is tedious, time-consuming and expensive. Computers Accuracy value of our model are 1.2% and 1.5% higher
have inherent advantages over human brains in terms of sample than VGG16, respectively. The experimental results indicate
storage capacity, data processing capacity, and calculation that the approach based Resent 50 is effective for distinguish
speed. The computer learns the relevant characteristics of a cancer patche images and normal patche images.
certain specific corresponding through massive medical data. So, the main highlights of this article are summarized as
In recent years, computer assisted diagnosis has received much follows:
attention from researchers. Computer assisted diagnosis uses • According to our existing knowledge, our work
machine learning algorithms and deep learning networks to is a new application innovation to recognize the
enable intelligent systems to actively capture disease-related metastatic cancer with Resent deep learning method.
features when facing a set of medical images, and to automat- • We evaluate our Resent model and other convolu-
ically classify and quantify patient medical images. tional neural networks on Pcam datasets. Our exper-
Image classification is an essential task in the field of image iments indicates that our method achieves superior
visual processing. Specifically, image classification aims to performance for cancer image classification.
distinguish different types of targets according to the different • Through studying the relationship between epoch and
characteristics reflected in the image data information. Medical loss value, we obtain some constructive guidances for
images recognition is a difficult subtask. Traditional machine work future.

978-1-7281-8141-7/20/$31.00 ©2020 IEEE 1356

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on August 28,2024 at 08:30:05 UTC from IEEE Xplore. Restrictions apply.
Fig. 1. The histopathological scans of lymph node section. The first row is
positive samples and the second row is negative samples.

Fig. 2. The architecture of Residual Neural Network model from the paper
The remainder of this article is laid out as follows. Second of He et al. [7]
section shows some related work on using deep learning
methods for the cancer image classification. Then, third section
introduces our Resent model for our task. After that, we con- alleviate the problems of the gradient explosion and gradient
duct many experiments on our proposed model and other deep disappearance problems due to the special connection struc-
neural networks in fourth section. Eventually, we summarize ture. The ResNet model uses skip connections with identity
this article and discuss future work in fifth section. transformation. The specific model structure (a two layers
block) from paper “Deep Residual Learning for Image Recog-
II. RELATED WORK nition” is shown in Fig. 2. Let H(x) is the basic mapping to fit
In this part, we will introduce some related work and latest some stacked layers where x is the input matrix of the first of
research progress about the cancer image classification task. stacked layers. Assuming the network layer approximate the
Histopathological cancer image classification is a difficult task. residual function F(x):
Traditional machine learning methods need lots of prepro-
cessing works and extract manual features extraction which F (x) = H(x) − x (1)
is labor intensive [1], [2], [9]. For example, Velusamy et al. where x represents the input matrix of the layer. Now, the
[1] diagnosed the breast cancer using SVM (Support vector mapping H(x) equals F(x) + x.
machine) and feature selection. The two ways could approximate the required model expres-
Recently, deep learning methods can effectively capture sion function, but the difficulty of learning might be different.
specific features and produce low-dimensional representations That is, calculating H(x) is more difficult than calculating F(x).
from image data types [10], [11]. There is a lot of work about For each stacked layers, the ResNet model uses the idea of
using deep neural networks to solve medical image problems residual learning. The calculation process of the final output
[12], [13]. Wang et al. [3] proposed a deep learning-based of the structure in Fig. 2 is as follows:
method VGG to identify cancer metastases from the entire
slide image of breast. Matsuyama et al. [4] used a CNN and a y = F (x, {Wi }) + Ws x (2)
wavelet transform approach for lung CT images classification,
which can identify 4 categories including metastatic lung where y is the output vector of the layer. F(.) is the residual
cancer. Ebigbo et al. [5] detected esophageal adenocarcinoma function to be learned and x is the input. Then the Ws is a the
by using a deep convolutional neural net with a residual net linear projection, to changes the dimension of the input to be
architecture on the image data. Zhu et al. [6] distinguish same as that of the output. We use this previous classic model
between breast cancer molecular subtypes by deep learning to solve new problems. It is a new application innovation
models GoogleNet, VGG. to recognize the metastatic cancer with Resent deep learning
However, this study combine deep learning idea with the method.
Resent model to identify the digital pathology scan images
IV. EXPERIMENTS
of metastatic cancer. Deep Residual Learning is very helpful
for training deep neural network which has a large number of A. Experimental Data and Experimental Settings
layers. Our framework can detect and classify cancer with the To evaluate our approach comprehensively, we conduct lots
highly Auc-Roc score and Accuracy score. of experiments on the PCam PatchCamelyon (PatchCamelyon)
benchmark dataset. In this paper, our PCam dataset is slenderly
III. METHODOLOGY modified which does not contain duplicate images. Therefore,
In this section, we introduce the proposed metastatic cancer our PCam dataset has 220025 image samples, where the
image classification model based on Resent model. number of positive samples is 89117 (i.e. the cancer mage,
He et al. [7] proposed the ResNet (Residual Neural Net- the tag is 1) and the number of negative samples is 130908
work) model. The ResNet could train a deeper network and negative samples (i.e. the normal image, the label is 0).

1357
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on August 28,2024 at 08:30:05 UTC from IEEE Xplore. Restrictions apply.
 In the experiment phase, we randomly divided the PCam As the epoch increases, the smaller the loss, the better the
dataset into the test dataset and training dataset according to model. But, as the epoch increases, the higher the accuracy, the
the ratio of 2:8. The size of the input matrix of each image better the model. In Fig. 3, when the number of epoch exceeds
sample is MxN. During training phase, the learning ratio is 2, the training loss and validation loss both keep decreasing.
0.0001 and the batch size is 64. To minimize the loss function, It indicate the performance of our model becomes better.
we use the Adam optimizer [14]. Then, we use a GPU to But, when the number of epoch is more than 8, two curves
implement our model by Pytorch. become flat. This shows the performance of the model keep
In this paper, the Auc-Roc score and the Accuracy score stable after getting a certain epoch. And more epochs will not
are the evaluation indicators to assess the performance of our increase the effect and will waste calculation costs. The similar
method and comparative methods. The Roc (receiver operating conclusions could also be obtained in Fig. 4. At the beginning,
characteristic) curve is used to visualize the performance of with the abscissa epoch grows, the ordinate accuracy shows
the classifier, and each point on the Roc curve reflects the a significant downward trend in both the training phase and
susceptibility to the same signal stimulus. So, the final model the verification phase. However, when the abscissa epoch
is selected based on the performance of the classifier on the becomes very large, the ordinate accuracy remains unchanged.
ROC. The Auc Roc score (between 0 and 1) meas the area Therefore, above results are constructive for guiding the future
under the Roc curve which is a tool used to measure the work in model training stage and verification stage. Choosing
imbalance in classification. The Auc Roc score can provide a a proper epoch is very important for model training and
comprehensive assessment of model performance. The larger validation.
the Auc-Roc score, the better the performance of the model.
Draw the Roc curve by plotting the false positive rate and the TABLE I
THE EXPERIMENTAL RESULTS OF OUR MODEL AND
true positive rate in different threshold settings. The formulas COMPARISON MODEL MODELS ON PCAM DATASETS
of true positive rate and false positive rate are calculated as
follows: Models Auc Roc Score Accuracy
Vgg16 0.951 0.957
Vgg19 0.955 0.962
# True positive ResNet50 0.963 0.972
True positive rate = (3)
#True positive + #False negative

# False positive
False positive rate = (4)
#False positive + #True negative
where # true positive represents the number of positive
samples which are correctly classified, and false negative
means the number of positive samples which are incorrectly
recognized. In the same way, false positive is the number
of negative samples which are falsely predicted, and true
negative denotes the number of negative samples which are
truly recognized. The Accuracy score is the sum of the four
above.
B. Experimental Results and Analyses
To evaluate the performance of our method, we conduct
experiments using our model and comparative model on the Fig. 3. The change of loss value under different epoch values in the training
same dataset and metrics. Table 1 shows the experimental stage and validation stage.
results of competing models and our model. It is easily known
that the Accuracy value of our ResNet50 model is 1.0% higher
than VGG19 from tabel 1. Auc Roc score is to compute area
under the receiver operating characteristic curve (ROC AUC)
from prediction scores, which reflects the pros and cons of
the algorithm. The Auc Roc score and the Accuracy value of
our model are 1.2% and 1.5% higher than VGG16 model.
The experimental results above indicate that applying the
Resent50 model is effective for our metastatic cancer image
classification task.
Fig. 3 draws the relationship between the epoch and the
loss value in the training stage and validation stage. And Fig. 4
shows the line chart between the epoch and the accuracy value.

1358
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on August 28,2024 at 08:30:05 UTC from IEEE Xplore. Restrictions apply.
 [7] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in Proceedings of the IEEE conference on computer vision
and pattern recognition, 2016, pp. 770–778.
[8] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[9] P. D. Velusamy and P. Karandharaj, “Medical image processing schemes
for cancer detection: A survey,” in 2014 International Conference on
Green Computing Communication and Electrical Engineering (ICGC-
CEE). IEEE, 2014, pp. 1–6.
[10] B. S. Veeling, J. Linmans, J. Winkens, T. Cohen, and M. Welling, “Ro-
tation equivariant cnns for digital pathology,” in International Confer-
ence on Medical image computing and computer-assisted intervention.
Springer, 2018, pp. 210–218.
[11] B. E. Bejnordi, M. Veta, P. J. Van Diest, B. Van Ginneken, N. Karsse-
meijer, G. Litjens, J. A. Van Der Laak, M. Hermsen, Q. F. Manson,
M. Balkenhol et al., “Diagnostic assessment of deep learning algorithms
for detection of lymph node metastases in women with breast cancer,”
Jama, vol. 318, no. 22, pp. 2199–2210, 2017.
[12] V. Sapra and M. Lal Saini, “Deep learning model for detection of
Fig. 4. The relationship between the accuracy value and the epoch in the breast cancer,” in International Conference on Advances in Engineer-
training stage and validation stage. ing Science Management & Technology (ICAESMT)-2019, Uttaranchal
University, Dehradun, India, 2019.
[13] M. Aubreville, C. Knipfer, N. Oetter, C. Jaremenko, E. Rodner, J. Den-
zler, C. Bohr, H. Neumann, F. Stelzle, and A. Maier, “Automatic
V. CONCLUSION classification of cancerous tissue in laserendomicroscopy images of the
oral cavity using deep learning,” Scientific reports, vol. 7, no. 1, pp.
Medical images recognition is a classification task in com- 1–10, 2017.
puter vision field. In this paper, we sovle the metastatic [14] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
detection task as a image binary classification task by using arXiv preprint arXiv:1412.6980, 2014.
ResNet model. The experimental results show that our method
outperforms than Vgg16 model and Vgg19 model. The Auc
Roc score and the Accuracy value of our model are 1.2%
and 1.5% higher than VGG16 model. We also studied the
relationship between the loss value and the epoch to obtain
some important guidances. Experiment indicate that the our
method based on Resent is effectively for metastatic cancer
image classification. In the future, we will utilize our model
to other Medical images classification tasks and improve
performance.

ACKNOWLEDGEMENT
The author Wang Mingrui and Xuhui Gong sincerely thank
to the Kaggle competition platform.

R EFERENCES
[1] M. F. Akay, “Support vector machines combined with feature selection
for breast cancer diagnosis,” Expert systems with applications, vol. 36,
no. 2, pp. 3240–3247, 2009.
[2] P. D. Velusamy and P. Karandharaj, “Medical image processing schemes
for cancer detection: A survey,” in 2014 International Conference on
Green Computing Communication and Electrical Engineering (ICGC-
CEE). IEEE, 2014, pp. 1–6.
[3] D. Wang, A. Khosla, R. Gargeya, H. Irshad, and A. H. Beck,
“Deep learning for identifying metastatic breast cancer,” arXiv preprint
arXiv:1606.05718, 2016.
[4] E. Matsuyama, D.-Y. Tsai et al., “Automated classification of lung
diseases in computed tomography images using a wavelet based convolu-
tional neural network,” Journal of Biomedical Science and Engineering,
vol. 11, no. 10, p. 263, 2018.
[5] A. Ebigbo, R. Mendel, A. Probst, J. Manzeneder, L. A. de Souza Jr,
J. P. Papa, C. Palm, and H. Messmann, “Computer-aided diagnosis using
deep learning in the evaluation of early oesophageal adenocarcinoma,”
Gut, vol. 68, no. 7, pp. 1143–1145, 2019.
[6] Z. Zhu, E. Albadawy, A. Saha, J. Zhang, M. R. Harowicz, and M. A.
Mazurowski, “Deep learning for identifying radiogenomic associations
in breast cancer,” Computers in biology and medicine, vol. 109, pp.
85–90, 2019.

1359
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on August 28,2024 at 08:30:05 UTC from IEEE Xplore. Restrictions apply.