A Low Layered Deep Network Model for the Classification of Skin Dermoscopy Images

doi:10.21203/rs.3.rs-2750070/v1

Download PDF

Research Article

A Low Layered Deep Network Model for the Classification of Skin Dermoscopy Images

https://doi.org/10.21203/rs.3.rs-2750070/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

One of the deadliest forms of skin cancer is Melanoma. This is increasing gradually, especially in males and can be treated if diagnosed at early stages. Manual inspection involves critical inspection by dermatologists and the decision requires a lot of expertise and consumes lot of time, hence there is a need to develop automated lesion detection and classification system with image processing and deep learning methods. There are several image processing methods were proposed in the research however there is a scope to automate the entire with deep learning methods. This paper proposes low layered, high performance deep network architecture for the classification of skin lesion images and evaluates the performance of the network with standard datasets. The work also focuses on comparing the performance evaluation metrics against state of art networks and highlights the need for new customized networks.

Melanoma detection

Skin lesion classification

Deep learning methods

In recent times, melanoma is treated as one of the deadliest type of skin cancer, specially found in men rather than women. It was estimated that around 76 thousand people were diagnosed with melanoma in United States alone [1]. This type of skin cancer accounts for around 75% of deaths that are associated with skin cancer [2]. The term melanoma means “black tumor”, which grows quickly and has the chances of spreading to other organs very quickly. These tumors emerge from skin cells called melanocytes that produces black pigment on the skin and observed that most of them are black or brown in color. It was observed that around 30%-35% of melanomas begin the emergence from the existing moles however the rest of the percent may start from normal skin. Early detection of melanoma is classified as stage I and the survival rate after detection is around 96% [3]. With increasing number of melanoma cases, it is recommended to screen for melanoma at early stages. In order to reduce the cost of screening and readily available to common people there is a need to develop automated melanoma screening methods and algorithms. Figure 1, depicts the difference between the Melanoma and Non-melanoma skin Images. As it can be observed that melanoma Images do contains the abnormal portion of the skin in very dark color and have irregular shape but the Non-Melanoma patches contains reddish spot which is coverable on the skin portions and possess regular shape.

The most used tool of image analysis in diagnosis of melanoma is dermoscopy which is non-invasive and cost effective approach of detecting early stage skin cancer. This process of dermoscopic image analysis has been proven to be more useful in reducing the presumptions that are involved in diagnosis which has to be confirmed using biopsy of skin portions [4]. These dermoscopic images are formed by combining optical magnification with either cross polarized lighting or liquid immersion with low angle of light incidence. The use of dermoscopic images are used to derive the border structural information of the tumor region which are usually asymmetric and irregular in shape. These regions are usually blue white areas which have globules with typical pigments. There are several pre-processing algorithms that were applied to dermoscopic images that include normalizing or enhancing the image colors [5]. However, it was observed that only 48% of dermatologists use these dermoscopic images for skin cancer analysis.

The following are the research objectives of the paper:

To design a low layered customized deep network model for the classification of skin dermoscopic images

To find out the most suitable activation function that best suits for the application.

To train and validate the proposed network model on different datasets and compare the performance with state of art existing network models.

This paper, presents a fully automatic image classification deep network for differentiating cancerous and non-cancerous images. In particular, this paper incorporates a customized fully connected convolutional network which comprises of few layers, application specific and requires low processing time with high accuracy for performing classification. The network is tested on standard datasets and compared against state of art networks in terms of classification accuracy and number of layers.

Semantic based analysis for predicting the skin lesion images was discussed in [6]. In this paper, the authors have presented the ontology focusing on the concepts of semantic annotation. In this work, the authors have extracted low level features that describe skin lesion shape feature, texture feature and color feature. Later a machine learning classifier is used to classify and which are based in ABCD rule for making decision for skin lesion prediction process. Experiments were conducted with 206 skin lesion images and demonstrated the performance to be 76.9% accuracy.

Sameena [7] and others have designed a new automatic system for detecting the pigment network and also provides a difference between the typical and non-typical patterns. In this work, the authors have tested the proposed approach on standard datasets with 200 images and attained an average accuracy of 96.7%.

A new framework with ABCD rules for classifying the melanoma skin images were proposed by Rajesh [8]. In this work, a new feature extraction method has been adopted that uses symmetry for recognizing the pictures and detecting the border for identifying the image color and also the diameter of the tumor region. In this work, LBP was used to extract the texture features and introduced back propagation network for classifying the melanoma or non-melanoma type of skin cancerous images.

A new model of classifying the melanoma tumor is proposed by Fengying and others [9]. The work aims to classify whether the dermoscopic image is benign or malignant. There are 3 important stages in proposed model. Firstly, the skin lesion regions were extracted using neural networks which they termed to be as self-generating neural network. Secondly, the features that were extracted from the skin lesion regions and lastly they were classified using ensemble based classifier.

An expert based automatic system is proposed by Suleiman et.al [10], that aimed to detect the melanoma skin cancer from the plain pictures that were captured from the effected regions. In this work, the authors have applied ABCDE rule to detect to detect the melanoma images. In this work, the authors have performed segmentation process of an input melanoma image to extract skin lesion region using Grabcut Algorithm and later the features are extracted for the segmented portion. Then the images are classified into cancerous and non-cancerous with support vector machine classifier with Radial Basis Function (RBF) kernel. In this, work the authors have tested with 100 of malignant and other 100 of benign images.

In recent times, deep networks have achieved considerable performance in several image processing and computer applications. In [11] Liu et.al, proposed deep learning system that aims to classify and diagnosis 26 types of skin diseases which are pre-dominant in many of the adults. In this work, the authors have aim to classify 26-types based on small different conditions that includes dermatitis, dermatoses and pigmentary conditions. Finally, the proposed deep network supports multiple input images and its benefit is assessed.

In other work, Han et.al [12], has proposed a deep network that is trained with large number of clinical samples which are more than 2 lakhs that were aim to differentiate 134 skin disorders. In the above works the network models have achieved higher performance yet they are difficult to incorporate and requires lot of time for training the system. So from the literature survey it is evident that there is scope to develop a customized deep network which comprises of low number of layers and yield high performance. In this paper, this scope is considered as the objective and aims to develop a deep network that comprises of fully connected convolutional layer which attains high performance with low layers and processing time.

The proposed deep network model for skin cancer classification is depicted in Fig. 2, this network model is a 24-layered architecture. The input to the network is a true color image of dimension 224x224x3 which is followed with a stack of layers comprises of convolutional layer, Batch normalization, drop-out, fully connected, soft-max and classification layers. The clear input dimensions of each layer is depicted in Fig. 2.

The input images were resized to 224x224 such that the image dataset is compatible for other networks. This size adopted for speeding up the process of training and validation. These images are fed to the image input layer and then processed for first convolutional layer. The input dimension of the first two dimensional convolutional layer are 16 filters each of 3x3 size for which the output dimensions of the image is reduced to half of the dimension of the original image. After two consecutive convolutional layers’, batch normalization that aims to reduce the internal covariance shift which makes the training of network faster. Later max pooling is applied which is used to reduce the extracted data from the convolutional layers. The output of the max pooling and dropout layers are forwarded for the sequence of convolutional layers which are lastly connected to fully connected layer which categorizes the data into two classes Melanoma and Non-Melanoma. At last, classification layer is used to give the class number by computing the cross entropy of the given input samples.

To validate the performance of the proposed network model, experiments were conducted on skin image dataset of university of waterloo [13] and compared against the existing networks like Vgg19 [14], ResNet-50 [15]and ResNet-101. Below table-1 shows the training options of the network model.

Table 1

Proposed Network Model Training Options
S.No	Training parameter	Type
1	Solver	SDGM (stochastic Descent Gradient method)
2	Initial Learning Rate	0.01
3	Validation Frequency	10
4	Max Epochs	50
5	Batch size	64
6	Momentum	0.99
7	Gradient Threshold Method	L2-norm

The dataset consists of two categories Melanoma and Non-Melanoma from which the images are resized and trained with 80% and validated with 20%. It was observed that the average accuracy of the proposed network model yielded more than the State of Art Network Models. Apart from this, the proposed model is tested with Ham 10000 dataset available Harvard data verse [16]. This dataset consists of part 1 and part 2 each of 5000 images and all these images [17] include all major types of diagnostic categories of pigmented lesions. Actinic keratosis, intraepithelial carcinoma (akic), basal cell carcinoma (bcc) and others which are provided in the meta data file in the dataset [18].

Figure 4 shows the percentage of training samples for each category of skin cancer images. It can be shown that both the categories are fed almost equally to the proposed network model.

Table 2

Percentage of Accuracy with the proposed Network Model for Waterloo Dataset
S. No	Network Model	No. Of. Layers	Percentage of Accuracy
1	VGG19[14] (Existing)	47	68.2%
2	ResNet 50 [15] (Existing)	177	71.3%
3	ResNet 101 (Existing)	347	72.2%
4	Proposed Model	24	83.9%

Though the proposed network model contains very low number of layers of 24 which is depicted in Fig. 3, which is trained with melanoma and non-melanoma images in the propositions shown in Fig. 4 for which it yields higher percentage of accuracy of 83.9% as shown in Fig. 7. The performances attained with different networks and its comparison is clearly shown in Fig. 5. The qualitative analysis of the proposed model with respect to other state of art models are presented in Table2. This percentage of accuracy of classification is higher when compared against the state of art methods. This achieves the objective of the work of designing a low number layered network model with high accuracy. The training options are provided in the Table 1 and training progress is depicted in Fig. 6 from these figures it was observed that the higher batch size results in higher accuracy, however the training validation percent of datasets also impacts the percentage of accuracy. Stochastic gradient descent with momentum type of solver is used with 50 epochs each of 100 iterations is used to train the network. At soft-max layer the loss function of ‘weighted cross-entropy’ is used to classify the images.

Table 3

Percentage of Accuracy with the proposed Network Model for Ham1000 Dataset
S. No	Network Model	No. Of. Layers	Percentage of Accuracy
1	VGG-19 [14]	47	79.43%
2	ResNet-50 [15]	177	79.31%
3	ResNet-101	347	77.55%
4	Proposed Model	24	88.6%

Figure 8 and Table 3 depicts the performances that were attained with different networks and its comparison with other state of art network models from which it can be stated that the proposed model could able to attain a higher percentage of accuracy more than 11 percent and Fig. 9 depicts the performance analysis of the proposed network model for the classification of Melanoma for Ham 10000 dataset. The proposed network model attains 88.6% accuracy which is 9 ~ 11% superior than other conventional network models, which are clearly shown in bar and confusion matrix.

This work aims to propose a customized low number of layered based deep network model for the classification of skin cancer images. In this work two categories of skin images were considered, melanoma and Non-melanoma. The proposed model is validated with standard image datasets and compared the percentage of classification accuracy against the state of art network models. The proposed model is compared against the VGG-19, ResNet-50, ResNet-101 models and observed that the proposed model yields high accuracy with low number of layers for the prescribed application. The main advantage with the proposed network model is that it consumes less times for processing the training samples and validating the samples. This work provides a pave for newer network models that are to be implemented for the specified applications.

Conflict of interest

We the authors don’t have any conflict of interest.

Acknowledgement

Research discussed doesn’t get any specific grant from funding agencies

We hereby declare that the information given above is complete and true to the best of my knowledge.

Data Availability

The datasets used in this research work are publicly available for research purpose in [13] [16].

N. Howlader, A. M. Noone, M. Krapcho, J. Garshell, N. Neyman, S. F. Altekruse, C. L. Kosary, M. Yu, J. Ruhl, Z. Tatalovich, H. Cho, A. Mariotto, D. R. Lewis, H. S. Chen, E. J. Feuer, and K. A. Cronin, “SEER cancer statistics review, 1975-2010,” Nat. Cancer Inst., Bethesda, MD, USA, Tech. Rep., 2013
A. F. Jerants, J. T. Johnson, C. D. Sheridan, and T. J. Caffrey, “Early detectionand treatment of skin cancer,” Amer. Family Phys., vol. 62, no. 2,pp. 1–6, Jul. 2000.
J. Glaister, A. Wong and D. A. Clausi, "Segmentation of Skin Lesions from Digital Images Using Joint Statistical Texture Distinctiveness," in IEEE Transactions on Biomedical Engineering, vol. 61, no. 4, pp. 1220-1230, April 2014, doi: 10.1109/TBME.2013.2297622.
L. Thomas and S. Puig, “Dermoscopy, digital dermoscopy and otherdiagnostic tools in the early detection of melanoma and follow-up ofhigh-risk skin cancer patients,” Acta DermVenereol, 2017
H. Iyatomi, M. Celebi, G. Schaefer, and M. Tanaka, “Automated color calibration method for dermoscopy images,” Comput.Med. Imag. Graph.,vol. 35, no. 2, pp. 89–98, Mar. 2011.
Wiem Abbes &DorraSellami , “Automatic Skin LesionsClassification Using Ontology-Based Semantic Analysis of OpticalStandard Images”, Procedia Computer Science, vol. 112, pp. 2096- 2105,2017
Sameena Pathan, K, Gopalakrishna Prabhu &P.C.Siddalingaswamy, “A methodological approach to classify typical and a typical pigment network patterns for melanoma diagnosis”, Biomedical Signal Processing and Control, vol. 44, pp. 25-37,2018
Rajesh, “Classification of malignant melanoma and Benign SkinLesion by using back propagation neural network and ABCD rule”,2017 IEEE International Conference on Electrical, Instrumentation and Communication Engineering (ICEICE), 2017
FengyingXie, Haidi Fan, Yang Li, Zhiguo Jiang, Rusong Meng & Alan Bovik, ‘Melanoma Classification on Dermoscopy ImagesUsing a Neural Network Ensemble Model’, IEEE Transactions onMedical Imaging, vol. 36, no. 3, pp. 849-858,2017
Suleiman Mustafa and Akio Kimura, “A SVM-based diagnosis ofmelanoma using only useful image features”, 2018 InternationalWorkshop on Advanced Image Technology (IWAIT).
Y. Liu, A. Jain, C. Eng, D. H.Way, K. Lee, P. Bui, K. Kanada, G. de OliveiraMarinho, J. Gallegos, S. Gabriele, and V. Gupta, “A deep learning systemfor differential diagnosis of skin diseases,'' Nature Med., vol. 26, no. 6,pp. 900-908, 2020.
S. S. Han, I. Park, S. EunChang,W. Lim, M. S. Kim, G. H. Park, J. B. Chae,C. H. Huh, and J.-I. Na, “Augmented intelligence dermatology: Deep neural networks empower medical professionals in diagnosing skin cancer and predicting treatment options for 134 skin disorders '', J. Investigative Dermatology, vol. 140, no. 9, pp. 1753-1761, Sep. 2020.
https://uwaterloo.ca/vision-image-processing-lab/research-demos/skin-cancer-detection
Nour Abuared; AlavikunhuPanthakkan; Mina Al-Saad; Saad Ali Amin; Wathiq Mansoor, “Skin Cancer Classification Model Based on VGG 19 and Transfer Learning”, 3^rd international conference on signal processing and information security, 2020
Mehra, A., Bhati, A., Kumar, A., Malhotra, R. (2021). Skin Cancer Classification Through Transfer Learning Using ResNet-50. In: Hassanien, A.E., Bhattacharyya, S., Chakrabati, S., Bhattacharya, A., Dutta, S. (eds) Emerging Technologies in Data Mining and Information Security. Advances in Intelligent Systems and Computing, vol 1300. Springer, Singapore.
https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/DBW86T
Salman M.N., Trinatha Rao P., Ur Rahman M.Z, ‘Adaptive noise cancellers for cardiac signal enhancement for IOT based health care systems’, Journal of Theoretical and Applied Information Technology,vol.95,no.10,PP.2206- 2213, 2017.
Tschandl, P., Rosendahl, C. &Kittler, H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Sci. Data 5, 180161 (2018). doi: 10.1038/sdata.2018.161

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

A Low Layered Deep Network Model for the Classification of Skin Dermoscopy Images

Status:

Version 1

Abstract

Figures

1. Introduction

2. Related Work

3. Proposed Network Model

4. Experimental Results

5. Conclusions

Declarations

References

Additional Declarations

Status:

Version 1