A Research Paper on Skin Cancer Detection Using Cnn

Abstract

This paper addresses the demand for an intelligent and rapid classification system of skin cancer using contemporary highly-efficient deep convolutional neural network. CNNs use convolutional layers to perform image processing on input images and learn to perform classification tasks. In order to take advantage of a tested and trusted network, the Alex Net CNN has been adopted. The developed program can be run by interacting with controls on its graphical user interface. There are two modes of operation. The first is to train the network with a dataset of choice while the second is to use the already trained network to classify the image via transfer learning.

Introduction

Advancements in digital technology, image processing, machine learning and more recently, deep learning has evolved the use of images for medical diagnoses. This trend is becoming popular with the ever-expanding number of machine learning methods and modern increasing computational power. As of now, several of deep learning models have been established or created which are applied in the field of medical diagnosis due to their ability and capacity to recognise patterns in digital images (Cicero et al., 2016).

Among the several deep learning techniques available, the convolutional neural networks (CNNs) are at this moment the best. Convolutional neural networks have led to major progress in many medical image analysis tasks such as detecting diseases and disease classification. One of the major areas of application of this technique in medicine is in the detection and classification of malignant and benign skin lesions from dermoscopy images. Other deep learning technique include deep neural networks, deep belief networks, recurrent neural networks and deep Boltzmann machine.

Eldeib, (2016) stated that an accurate CAD system can be used for early detection of a disease and thereby allow for earlier and more effective treatment which could saves lives. For example, the ability to efficiently and accurately cure and treat cancer is variably dependent on the ability to detect cancers at their earliest stages.

According to Choi et al. (2010), cancer has been classified and differentiated from diseases that are related where diagnosis and treatment are of high interest due to high rate of occurrence. Cancer has been identified as a common cause of death for humans as records show there were 14 million new cases of cancer and 8.2 million cancer related deaths worldwide (National Cancer Institute, 2017). Due to this, CAD has the eligibility and potential to use only images of the skin lesion without any other information because skin cancer is one of the most common types of cancers and it’s usually forms in the skin that has been vulnerable to sunlight.

Literature survey

A. Classification on the basis of types

In this study, the focus is basically on three main types of skin cancer, which are firstly, naevus, a skin cancer group that consists of various kinds of dysplastic naevus, melanocytic naevus, epidermal naevus, etc. Secondly, we have the seborrheic keratosis class which are mainly benign and thirdly we have the malignant and highly dangerous melanoma.

1. Naevus

The name naevus or nevi in Latin means birthmark. This skin cancer is usually benign and may be like melanoma in appearance. People who develop naevus are highly prone to developing melanoma in a mole or elsewhere on the body. The higher the number of these moles acquired, the greater the risk. The people having ten or more, have twelve times the risk of developing melanoma compared to the general population.

2. Seborrheic Keratosis

These are generally harmless skin growths that often appear as the skin ages. Some people have just one, but it is not uncommon to develop several more. This cancer does not present any risk to a patient. They are often brown and patchy and can appear anywhere on the body. The growth may look rigid as if they are painted onto the body. Some people mistake them for unusual looking scabs. It occasionally causes pain and itching which later become inflamed or appalling.

3. Melanoma

This arise the least frequent but most deadly of the three common skin cancers. They spread rapidly and can potentially result in death. It is often difficult to differentiate between a naevus skin cancer and an early melanoma. Certain features of a naevus skin mole can determine whether it’s friendly or at moderate or high risk of becoming melanoma. The traditional way of detecting this cancer is using a dermoscope which a handheld is, high tech magnifying device that allows the visualization of internal skin structures and colour that are not visible to the eye.

B. Skin Lesions Classification Based On CNN

There are two ways in which a CNN can be used to classify skin cancer. The first is to use a pretrained CNN such as Alex Net, Google Net or Image Net which has been trained and configured with a dataset containing millions of images (Hinton et al., 2012). The CNN is retrained using a fresh dataset containing skin cancer images. The CNN completes its function as only a feature extractor. Classification is performed by using another classifier, such as support vector machines or artificial neural networks.

The second way, the one employed in this study is to make the CNN process the images from a dataset, extract and learn features that are peculiar to the different categories of the dataset. The CNN directly learns the relationship between raw pixel data and the class labels through end-to-end learning. Feature learning is integrated into the workflow and is not separate from classification. In other to develop a fully functional classification computer program based on CNN, the process is separated into two. The first is to train the CNN which in our case is Alex Net. After training the CNN is saved for future use. Secondly, the saved CNN is used to classify images through the use of transfer learning.

The accuracy of CNN training is only as good as its dataset. To successful train deep CNN models, all images contained in the dataset must be properly labelled. The labels are based on the categories available for classification. If this is done wrongly, it can result in overfitting the network which will lead to improper generalization in the network for unfamiliar input images.

Statement of problem

Presently, DERMOSCOPY is used as a tool to specifically detect the skin lesions of melanoma, because it can be diagnosed in its early stage and there is assurance that a good chance of recovery is certain. However, the diagnosis through dermoscopy images is quite difficult as it requires specialist that are extensively and extremely trained. Because of this, deep learning models such as CNNs are currently deployed for automatic detection of melanoma from dermoscopy images.

These existing CNNs architectures have used several types of convolutional architectures and various ways for classifying their prediction accuracy. Therefore, the challenges being studied in this work is to create a convolutional architecture in such a way that it can suck useful properties or features from the bio-medical images for high accuracy melanoma classification. This paper proposes a model based

Methodology

A. Dataset Used

The data used in this study are a combination of datasets from the following sources which include the International Skin Imaging Collaboration (ISIC, 2018), MED-NODE database (Giotis et al., 2015) and DERMOFIT image library. The images are dermoscopy images that are publicly available in from these sources. They are stored as RGB images with varied sizes in JPG format. There were total 1609 images in the dataset. The images were divided into two parts. These are 90% and 10% respectively. The first part was used for the training of the CNN while the second part was used validation.

B. CNN Used - AlexNET

AlexNet is a convolutional neural network that is trained on more than a million images from the ImageNet database. The network is 8 layers deep and can classify images into 1000 object categories, such as keyboard, mouse, pencil, and many animals. As a result, the network has learned rich feature representations for a wide range of images. The network has an image input size of 227-by-227

Syntax net = alexnet returns a pretrained AlexNet network.

C. Evaluation Metrics

In order to quantify the reliability of the trained convolutional network, two important factors are calculated. They include the sensitivity and classification accuracy. These are based on validation loss, all cases, true detected cases, true positive cases and false negative cases.

Sensitivity

sensitivity=True positive casesTrue positive cases+False negative cases

Accuracy

Accuracy=True detected casesAll cases

Results

The dataset used for this study contains 545 melanoma images, 651 naevus images and 413 seborrheic keratosis images. The average training time is 40 min and 18 sec. Also, all training sessions passed the validation criteria of completing training before the last epoch. A maximum of 6 epoch was used for training.

For the system configuration, the minimum batch size is 10, initial learning rate is1e-6, the input data is shuffled after every epoch, validation of learning progress to determine learning accuracy is performed after every 10 iterations. Verbose is set to false, so there is no logging information. On completion of training, the network is saved to be utilized for image classification via transfer learning.

When the convolutional network training starts, the accuracy of the training is just above 20% while the losses are as high as 1.5. As training progresses and the number of iterations increases, the accuracy increases rapidly and slows down towards the completion of epoch 1. The losses are inversely related to training accuracy. With increasing accuracy, the value of losses decreases. With increasing epoch number, the training slows, and increases become gradual. It finally approaches constant training accuracy values. In the training progress result presented in Figure, the constant value for training accuracy progress reached during epoch 6 at the 800th iteration.

An epoch is a complete set of input data feed into the convolutional network that is to be learned. Multiple epochs are needed by the deep learning algorithm during their learning phase to achieve high classification accuracy. An epoch in our experiment consists of 152 iterations. In all there is a maximum of 912 iterations. An iteration means the updating of the deep learning algorithm.

After the network has been trained, it is tested with validation images and some of them are displayed on a result window.

From tests conducted, the maximum accuracy of the network is 74% while a minimum of 68% was achieved. Also, sensitivity analysis was carried out and was as high as 70%. This is a reflection of the fact that the program’s potential of achieving almost a 100% accuracy is very high if the proper variables are optimized. It should however be noted that optimization takes both time and optimization power. The dataset size will be increased along with the batch size and number of epoch. Both underfitting and overfitting of the data will be totally eliminated.

Conclusion

The use of deep learning in the field of medicine for medical diagnoses through dermoscopic images have been identified as a future dominant technology due to its potential for high accuracy reduction of resources in terms of time and cost. With a large number of medical image dataset available for public use it is essential to develop tools to be able to fully utilize these resources.

The program was developed in MATLAB and one of its special features is a graphical user interface that makes it easy to use. The deep learning program developed in this study is capable of the following:

1. Record results on highly challenging datasets
2. The program can be retrained severally
3. Its program is easy to use.

The program promises to provide a viable alternative to skin cancer classification.

Dataset quality determines the training result. From the experiments carried out, the highest accuracy achieved was 74 % while the lowest was 68 %. Sensitivity is as high as 70 %. Further development of this program will ensure the accuracy is close to 100% and it can serve as a basis for professional skin cancer classification software in the medical industry.

References

• Abbas, Q., Emre Celebi, M., Garcia, I. F., & Ahmad, W. (2013). Melanoma recognition framework based on expert definition of ABCD for dermoscopic images. Skin Research and Technology, 19(1), e93–e102. https://doi.org/10.1111/j.1600-0846.2012.00614.x
• Abdel-Zaher AM, Eldeib AM. Breast cancer classification using deep belief networks. Expert Systems with Applications. 2016;46:139–144.
• Barata, C., Ruela, M., Francisco, M., Mendonça, T., & Marques, J. S. (2014). Two systems for the detection of melanomas in dermoscopy images using texture and color features. IEEE Systems Journal, 8(3), 965–979.
• C. T. R. Kathirvel. Classifying Diabetic Retinopathy using Deep Learning Architecture. International Journal of Engineering Research Technology, 5(6), 2016.
• Cavalcanti, P. G., Scharcanski, J., & Baranoski, G. V. G. (2013). A two-stage approach for discriminating melanocytic skin lesions using standard cameras. Expert Systems with Applications, 40(10), 4054–4064. https://doi.org/10.1016/j.eswa.2013.01.002
• Chang, W.-Y., Huang, A., Yang, C.-Y., Lee, C.-H., Chen, Y.-C., Wu, T.-Y., & Chen, G.-S. (2013). Computer-aided diagnosis of skin lesions using conventional digital photography: a reliability and feasibility study. PloS One, 8(11), e76212.
• Choi YE, Kwak JW, Park JW. Nanotechnology for early cancer detection. Sensors. 2010;10(1):428–455.
• Cicero, F., Oliveira, A., Botelho, G., & da Computaçao, C. de C. (2016). Deep learning and convolutional neural networks in the aid of the classification of melanoma. SIBGRAPI. Retrieved: http://sibgrapi.sid.inpe.br/col/sid.inpe.br/sibgrapi/2016/09.01.15.04/doc/16.pdf
• Codella, N., Nguyen, Q.-B., Pankanti, S., Gutman, D., Helba, B., Halpern, A., & Smith, J. R. (2016). Deep Learning Ensembles for Melanoma Recognition in Dermoscopy Images. arXiv:1610.04662 [Cs]. Retrieved from http://arxiv.org/abs/1610.04662
• Codella, N., Cai, J., Abedini, M., Garnavi, R., Halpern, A., & Smith, J. R. (2015). Deep Learning, Sparse Coding, and SVM for Melanoma Recognition in Dermoscopy Images. In Machine Learning in Medical Imaging (pp. 118–126). Springer, Cham. https://doi.org/10.1007/978-3-319-24888-2_15
• Doi W, Yang Q, Xue GR, Yu Y. Boosting for transfer learning. In: Proceedings of the 24th international conference on machine learning. ICML ’07. ACM; 2007. p. 193–200.
• Dumoulin, V., & Visin, F. (2016). A guide to convolution arithmetic for deep learning. arXiv Preprint arXiv:1603.07285. Retrieved from https://arxiv.org/abs/1603.07285
• Egmont-Petersen, M., de Ridder, D., & Handels, H. (2002). Image processing with neural networks—a review. Pattern Recognition, 35(10), 2279–2301. https://doi.org/10.1016/S0031-3203(01)00178-9
• Fornaciali, M., Carvalho, M., Bittencourt, F. V., Avila, S., & Valle, E. (2016). Towards automated melanoma screening: Proper computer vision & reliable results. arXiv Preprint arXiv:1604.04024. Retrieved from https://arxiv.org/abs/1604.04024
• G Wimmer, S Hegenbart, A Vecsei, and A Uhl. Convolutional neural network architectures for the automated diagnosis of celiac disease. In International Workshop on Computer-Assisted and Robotic Endoscopy, pages 104–113. Springer, 2016.
• Gola Isasi, A., García Zapirain, B., & Méndez Zorrilla, A. (2011). Melanomas non-invasive diagnosis application based on the ABCD rule and pattern recognition image processing algorithms. Computers in Biology and Medicine, 41(9), 742–755. https://doi.org/10.1016/j.compbiomed.2011.06.010
• Harry Pratt, Frans Coenen, Deborah M Broadbent, Simon P Harding, and Yalin Zheng. Convolutional neural networks for diabetic retinopathy. Procedia Computer Science, 90:200–205, 2016.
• He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2016 Presented at: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); June 27-30, 2016; Las Vegas, NV.
• Pelmer M Wolterink, Tim Leiner, Max A Viergever, and Ivana Iˇsgum. Automatic coronary calcium scoring in cardiac ct angiography using convolutional neural networks. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 589–596. Springer, 2015.
• Konstantinos Kamnitsas, Christian Ledig, Virginia FJ Newcombe, Joanna P Simpson, Andrew D Kane, David K Menon, Daniel Rueckert, and Ben Glocker. Efficient multi-scale 3d cnn with fully connected crf for accurate brain lesion segmentation. Medical Image Analysis, 36:61–78, 2017
• Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet Classification with Deep Convolutional Neural Networks. In F. Pereira, C. J. C. Burges, L. Bottou, & K. Q.
• Weinberger (Eds.), Advances in Neural Information Processing Systems 25 (pp. 1097– 1105).Curran Associates, Inc. Retrieved from http://papers.nips.cc/paper/4824-imagenetclassification-with-deep-convolutional-neural-networks.pdf
• Marín, C., Alférez, G. H., Córdova, J., & González, V. (2015). Detection of melanoma through image recognition and artificial neural networks. In World Congress on Medical Physics and Biomedical Engineering, June 7-12, 2015, Toronto, Canada (pp. 832–835). Springer. Retrieved from http://link.springer.com/chapter/10.1007/978-3-319-19387-8_204
• Mengqi Pei, Xing Wu, Yike Guo, and Hamido Fujita. Small bowel motil-ity assessment based on fully convolutional networks and long short-term memory. Knowledge-Based Systems, 121:163–172, 2017.
• National Cancer Institute. What Is Cancer?; 2015. Accessed: 2018- 03-03. Available from: https://www.cancer.gov/aboutcancer/ understanding/what-is-cancer#tissuechanges- not-cancer.
• Nima Tajbakhsh, Suryakanth R Gurudu, and Jianming Liang. Automatic polyp detection in colonoscopy videos using an ensemble of con-volutional neural networks. In Biomedical Imaging (ISBI), 2015 IEEE 12th International Symposium on, pages 79–83. IEEE, 2015.
• Poudel RPK, Lamata P, Montana G. Recurrent fully convolutional neural networks for multi-slice MRI cardiac segmentation. Vol. 10129 of Lect Notes Comput Sci 2017:83–94.
• Rongsheng Zhu, Rong Zhang, and Dixiu Xue. Lesion detection of endoscopy images based on convolutional neural network features. In Image and Signal Processing (CISP), 2015 8th International Congress on, pages 372–376. IEEE, 2015.
• The Skin Cancer Foundation. Skin Cancer Information; 2018. Accessed: 2018-04-25. Available from: https://www. skincancer.org/skin-cancer-information