A novel and efficient approach for the classification of skin melanoma

Abstract

Melanoma is a very common type of skin cancer that is an out-of-control growth of the cell. The detection of melanoma is very crucial in clinical practice due to irregularity in the shape of border and difference of tissue. Even though there are several methods proposed in literature, the identification of melanoma is still a complex process. This work proposes a novel Color Local Directional Pattern-based feature extraction technique for the classification of the skin cancer images. Along with the Color Local Directional Pattern this work also uses various descriptors for shape, color and Pyramid Histogram of Oriented Gradients. This method also uses a stacked Restricted Boltzmann Machine in Deep Belief Network, Support Vector Machine and Random Forest classifier for the classification. The experiments used the images from International Skin Imaging Collaboration (ISIC) 2016, ISIC 2017, Pedro Hispano Hospital (PH2) dataset, Dermnet and DermIS datasets. The performance metrics like specificity, sensitivity, F-Score, positive predictive value, accuracy, harmonic mean, and Area Under the Curve (AUC) are used for the experimental analysis. The results show that the proposed work achieves best results while using the combination of texture, color and shape descriptors and stacked RBM in DBN.

Appendix 1: To calculate the CLDP feature vector using the proposed method, perform the following steps. Consider a 3 × 3 block in the given image as shown below.

Step 1: Obtain R, G, B color channel images from the preprocessed sample image in dataset. A sample 3 × 3 block from each channel is given below.

Step 2: Create combined channel images like RR, GG, BB, RG, RB and GB. Convolution of each 3 × 3 block is done with the 3 × 3 eight directional Kirsch masks.

CLDP_R,R(u,v,θ) CLDP_G,G(u,v,θ) CLDP_B,B(u,v,θ) CLDP_R,G(u,v,θ) CLDP_R,B(u,v,θ) CLDP_G,B(u,v,θ).

The CLDP images are created for different channel combinations as in Eq. (1). The $i, j$ values used in Eq. (1) are determined according to the angular positions defined in Eq. (2). The Kirsch compass masks are used for convolution with the combined channel images. The Kirsch masks are given in Fig. 11

Step 3: Creation of feature vectors.

Consider the Kirsch mask responses obtained after convolution, form six CLDP code images using only the maximum responses to eliminate the random noise included in images during image acquisition. Then, the feature vectors are created from the histograms of the six CLDP code images. The $\mathrm{code }\left(\mathrm{x},\mathrm{y}\right)$ is created as given in Eq. (3) by considering the maximum of responses obtained from the eight directional masks given in Fig. 9. From the code(x,y) obtained for each pixel as in Fig. 9 the CLDP code image is formed. The feature vector of each code image is calculated using its histograms obtained from the grids as in Eq. (4) and as in Fig. 10.

Step 4: Concatenate all the feature vectors.

As in Eq. (5) all the feature vectors are concatenated to obtain the final feature vector used for classification.

The feature vectors thus obtained helps in extracting the color, texture, edge and structural information from the images excluding noise which helps in obtaining good classification accuracy.