Hybrid Melanoma Detection Using Neural Network Classifier

Abstract

Abstract—In spite of the gargantuan number of patients affected by melanoma every year, its detection at an early stage is still a challenging task. This paper illustrates a method which involves the combination of the existing ABCD (Involving symmetry, border, color, and diameter detections) rule and Gray Level Co-occurrence Matrix (GLCM) along with Local Binary Pattern (LBP) to identify the malignant melanoma skin lesion with greater accuracy. Several steps such as image acquisition technique, pre-processing (RGB to HSV), and segmentation are undertaken for the skin feature selection criteria to successfully determine the lesion characteristics for classification. Texture features such as contrast, energy, entropy, and homogeneity of the skin lesion is extracted using the LBP and GLCM for discriminatory purposes of the two cases (melanoma and non-melanoma). Back Propagation Neural Network (BPN) is used for the process of classification.

Introduction

Cancer is an abnormal growth of cells which tend to proliferate in an uncontrolled way and, in some cases, to metastasize. It can involve any tissue of the body and have many different forms in each body area. It is a group of more than 100 different and distinctive diseases of varying severity. Most cancers are named for the type of cell or organ in which they start. Out of the various types of cancer, skin cancer is one among the most common. The annual cost of treating skin cancers in the U.S. is estimated at $8.1 billion, of which about $4.8 billion is for nonmelanoma skin cancers and $3.3 billion is for melanoma. As per the studies released by The American Cancer Society in 2018, at least one instance of death is due to melanoma every hour. Despite the instance of high mortality, if diagnosed and detected at an early stage, it can be treated. Early detection techniques involve dermoscopic techniques and image processing techniques such as ABCD rule, pattern feature analysis, and many more. Despite such existing methods the diagnosis of the melanoma at an earlier stage with accuracy is a challenging task.

In the past, a mainly computer-aided pattern classification system for dermoscopy images was used which utilized a pattern classification system. This method was seen to have accuracy and several parametric values for the skin lesion but had a setback of high computational complexity. Later on a more unsupervised approach was desired for the detection. This led to the usage of SVM (Support Vector Machine) which involves a large number of training sets. In contrast to this ABCD rule is seen to have a much lower number of training sets and lower computational complexity if one takes and compares only its most best-performing score parameters into consideration.

Aside from the aforementioned approaches, melanoma detection using Delaunay triangulation is also a popular method which yields decent accuracy results if not for its slow convergence rate. Recently a detection methodology involving the Histogram of Oriented Gradient (HOG) and the Histogram of Oriented Lines (HOL) is employed which extracts the Bagged textural and color features of the skin lesion. This method has a major drawback that it can achieve good results only if there is a high contrast between the lesion areas. In comparison, Gray Level Co-occurrence Matrix (GLCM) coupled with Local Binary Pattern (LBP) extracts texture features of the skin lesions and uses them for the diagnosis of melanoma. It is seen to achieve good results irrespective of the contrast, unlike the histogram method.

Methodology

The existing methodologies only involve the ABCD parameters which is used for classification or the texture features extracted using GLCM and LBP methods. The proposed method is a combination of the ABCD rule and the GLCM coupled with LBP using a Back Propagation Neural Network as a final classifier. This approach is to minimize classification errors and boost the accuracy of the process.

Initially using the segmentation procedure the skin lesion is extracted by employing the adaptive thresholding algorithm. Then the feature extraction takes place giving the texture and color parameters of the image. The input dermoscopic image is subjected to pre-processing techniques such as color space conversion. This step is to extract the characteristic color features of the image subjected to observation. It involves the conversion to find the hue, saturation, and value (HSV) of the image. This gives information such as identity, purity, and intensity of the color.

A local binary pattern operator is outlined as a gray level invariant texture living in a local neighborhood. Then the Local Binary pattern operator labels the pixel of a picture by threshold the 3X3 neighborhood of every pixel and concatenating the results binomially to make a variety. LBP operator helps in classifying the image region as uniform or non-uniform. GLCM considers the relationship between two nearby pixels called reference and neighbor pixel. It is comprised of a matrix which has rows and columns equivalent to the number of gray levels of the image. By assigning the matrix elements with intensity and pixel distance factors various texture features can be computed. Out of the 14 texture features the best-performing texture features with the highest discriminatory factor is considered such as contrast, energy, homogeneity, and entropy which is extracted using GLCM and LBP. These parameters are compared with non-melanoma cases for comparison.

Melanoma & Non- Melanoma

Parallelly the features extracted are used for the ABCD classification which involves giving specific score values to characteristic features such as area, perimeter, major axis length, minor axis length, solidity, centroid, and orientation. These values which significantly depict the symmetry, border, color, and diameter of the skin lesion is used as a comparison between melanoma and non-melanoma cases.

The data from the ABCD methodology and the GLCM methodology is fed to the back propagation neural network which is used to act as a classifier. The experimental data comprises of 120 dermoscopic images of which 20 of them are testing images and 100 are training images. The training sample features with assigned target vectors are fed into the created BPN model for supervised training to get network parameters such as node biases and weighting factors. Finally, the test image features are simulating with trained network to make decision of skin lesion stages like malignant or benign.

Related Works

There are several systems for the identification of melanoma in dermoscopy images. The classification of the skin lesions is done in the global region extracted from the dermoscopy image. In the global method, the process of segmentation is done using a simple adaptive thresholding algorithm. GLCM matrix is used for extracting the texture features in four different orientation angles. There are several systems for the identification of melanoma in dermoscopy images. The classification of the skin lesion is done in the global region extracted from the dermoscopy image. In the global method, the process of segmentation is done using a simple adaptive thresholding algorithm. GLCM matrix is used for extracting the texture features in four different orientation angles. There are several systems for the identification of melanoma in dermoscopy images. The classification of skin lesions is done in the global region extracted from the dermoscopy image. In the global method, the process of segmentation is done using a simple adaptive thresholding algorithm. GLCM matrix is used for extracting the texture features in four different orientation angles. There are several systems for the identification of melanoma in dermoscopy images. The classification of the skin lesions is done in the global region extracted from the dermoscopy image. In the global method, the process of segmentation is done using a simple adaptive thresholding algorithm. GLCM matrix is used for extracting the texture Features in four different orientation angles.

There are several systems for the identification of melanoma in dermoscopy images. The classification of the skin lesion is done in the global region extracted from the dermoscopy image. In the global method, the process of segmentation is done using a simple adaptive thresholding algorithm. GLCM matrix is used for extracting the texture features in four different orientation angles.

In this, segmentation is done using a Gaussian filter and Otsu’s method is used to compute the global threshold. Feature extraction is done using 2D-fast Fourier transform, 2D-Discrete cosine transform, Pigment network feature, and color. The classification is done using SVM-RBF for the detection of melanoma.

This paper presents an automated method for melanoma diagnosis applied on a set of dermoscopy images. Features extracted are based on gray level Co-occurrence matrix (GLCM) and Using multilayer perception classifier (MLP) to classify between melanocytic Nevi and malignant melanoma. MLP classifier was proposed with two different techniques in the training and testing process: Automatic MLP and traditional MLP. Results indicated that texture analysis is a useful method for discriminating of melanocytic skin tumors with high accuracy.

In 2011, Daniel Ruiz, Vicente Berenguer, Antonio Soriano, and Belen Sanchez proposed types of ANN classifiers, which area multilayered perception, a Bayesian classifier, and the algorithm of the k nearest neighbors. These methods work independently and also in combination making a collaborative decision support system. The classification rates obtained are around 87%. An internet-based melanoma screening system was proposed in which ther server is opened for the public to upload the dermoscopy images, In this system, the digital dermoscopic image can be uploaded by the visitor and can register the clinical and pathological data. Once the image is accepted by the server, the tumor area is automatically extracted from the surrounding skin using an automatic threshold decision algorithm.

Experimental Result

This section consists of the results obtained from the classification of the images by means of the two employed dermoscopic methodologies. The experimental data comprises of 120 dermoscopy images, previously diagnosed, of which 20 of them are testing images and 100 training images. It contains images of 60 melanocytic nevi and 60 melanomas in total. After segmentation, the feature extraction involving the color and its texture is computed by the GLCM and LBP spatial relationship. Using GLCM and LBP a list of four texture parameters which performs the best are updated as weighted values in the table. Table 1 shows the results of both malignant and benign cases. The tabulation which contains a range of the parameters of both instances (melanoma and non-melanoma) is shown in table 2. The ABCD parameters consisting of area, major axis length, minor axis length, eccentricity, solidity, centroid, and orientation are also extracted and tabulated for both cases in table 3. A list portraying the comparison of the average GLCM parametric values obtained from other journals to the proposed methodology is shown in table 4. For the final classification, a BPN classifier is used with the weighted nodes as input. The proposed method is trained with 83.3% and tested with 16.7% of the total number of images.

References

1. Catarina Barata, Margarida Ruela, Mariana Francisco, Teresa Mendonca and Jorge S.Marques, “Two Systems for the Detection of Melanomas in Dermoscopy Images Using Texture and Color Features”, IEEE systems Journal, Vol.8,No.3, Sep.2014.
2. J.C. Kavitha , suruliandi.A , “Texture and color feature extraction for classification of Melanoma Using SVM”, ICCTIDE conference,jan 2016.
3. Omar A., Buket d.Barkana, Non invasive Real Time Automated skin Analysis system for Melanoma Early Detection and prevention”, IEEE Journal of Translation Engineering in Health and Medicine, April 2015.
4. Mariam A. sheha, Mai S.Mabrouk, Amr sharawy “Automatic Detection of Melanoma Skin Cancer using Texture Analysis” International journal of computer applications, vol.42,No.20, March 2012.
5. Daniel Ruiz, Vicente Berenguer, Antonio Soriano and Belen Sanchez “A decision support system for the diagnosis of melanoma.” June 2011.
6. H. Iyatomi, H. oka, M. E. Celebi, M. Hashimoto, M.Hagiwara,” An internet based melanoma Diagnosistic system- Towards the Practical Application” IEEE,2005
7. Catarina Barata, Margarida Ruela, Mariana Francisco “Two Systems for the Detection of Melanomas in Dermoscopy Images Using Texture and Color Features”, IEEE systems Journal, Vol.8,No.3, Sep.2014.
8. G.Di Leo, A.Paollillo, P.Sommella and C.Liguori, “ An improved procedure for the automatic detection of dermoscopic structures in digital elm images of skin lesions, in Proc.2008 IEEE computer soc.
9. Vishaka Sharma, “Melanoma Skin Cancer Detection Using Image Processing”, WCESC (2017), 03–09 pg
10. G.C do Carmo and M. R. e Silva, “Dermoscopy: basic concepts”,vol 47,pg: 712-719 (2008)
11. H. Kittler, H. Pehamberger, K. Wolff, “Diagnostic accuracy of dermoscopy”, Lancet Oncology, vol 3, pg:159-65, (2002)
12. Pehamberger H, Binder M, Steiner A, Wolff K. “In vivo epiluminescence microscopy: improvement of early
13. diagnosis of melanoma” J Invest Dermatol, pg: 56S–62S, (1993)
14. Demyanov S, Chakravorty R, Abedini M, Halpern A, Garnavi R. “Classification of dermoscopy patterns using deep convolutional neural networks”; 2016 IEEE.