Skin Lesion Classification Based on Surface Fractal Dimensions and Statistical Color Cluster Features Using an Ensemble of Machine Learning Techniques

Abstract

(1) Background: An approach for skin cancer recognition and classification by implementation of a novel combination of features and two classifiers, as an auxiliary diagnostic method, is proposed. (2) Methods: The predictions are made by k-nearest neighbor with a 5-fold cross validation algorithm and a neural network model to assist dermatologists in the diagnosis of cancerous skin lesions. As a main contribution, this work proposes a descriptor that combines skin surface fractal dimension and relevant color area features for skin lesion classification purposes. The surface fractal dimension is computed using a 2D generalization of Higuchi's method. A clustering method allows for the selection of the relevant color distribution in skin lesion images by determining the average percentage of color areas within the nevi and melanoma lesion areas. In a classification stage, the Higuchi fractal dimensions (HFDs) and the color features are classified, separately, using a kNN-CV algorithm. In addition, these features are prototypes for a Radial basis function neural network (RBFNN) classifier. The efficiency of our algorithms was verified by utilizing images belonging to the 7-Point, Med-Node, and PH2 databases; (3) Results: Experimental results show that the accuracy of the proposed RBFNN model in skin cancer classification is 95.42% for 7-Point, 94.71% for Med-Node, and 94.88% for PH2, which are all significantly better than that of the kNN algorithm. (4) Conclusions: 2D Higuchi's surface fractal features have not been previously used for skin lesion classification purpose. We used fractal features further correlated to color features to create a RBFNN classifier that provides high accuracies of classification.

Keywords: Higuchi fractal dimensions; Radial basis function neural network; k-nearest neighbor; skin cancer recognition.

Figure 1

Processing steps for classifying skin lesions.

Figure 2

Illustration of pre-processing for hair removal and segmentation. Rows 1 and 2: nevi image (PH2 database). Rows 3 and 4: melanoma image (7-Point database in row 3 and Med-Node database in row 4). First column: original image. Second column: image after the hair is removed. Third column: segmented image. Med-Node contains non-dermoscopic (or simple digital) images acquired via cross-polarized light. Skin lesions in PH2 were imaged via polarized noncontact dermoscopy by using conventional cross-polarized light. Skin lesions in 7-Point were imaged via fluid immersion and non-polarized dermoscopy.

Figure 3

Example of measuring the color content and clustering method to select the relevant color distribution in melanocytic lesion images. 23 color clusters, denoted as cl1, …, cl23, are presented. Data for the minimum and maximum intensity in each R, G, and B channel are indicated for each color cluster. Lower right: the segmented image.

Figure 3

Example of measuring the color content and clustering method to select the relevant color distribution in melanocytic lesion images. 23 color clusters, denoted as cl1, …, cl23, are presented. Data for the minimum and maximum intensity in each R, G, and B channel are indicated for each color cluster. Lower right: the segmented image.

Figure 3

Example of measuring the color content and clustering method to select the relevant color distribution in melanocytic lesion images. 23 color clusters, denoted as cl1, …, cl23, are presented. Data for the minimum and maximum intensity in each R, G, and B channel are indicated for each color cluster. Lower right: the segmented image.

Figure 4

An example of the HFD computation. (a–c) R, G, and B color channels for a digital image that belongs to the PH2 dataset. (d–i) An illustration of the tessellation patterns for each color channel. The triangle shapes of k = 1 are presented in (d–f); the fractal scaling parameter k = 4 is shown in (g,i). They are used to compute the $$X_{k_c}^m$$ matrices. (d,g) are R channel images. (e,h) are G channel images. (f,i) are B channel images.

Figure 4

An example of the HFD computation. (a–c) R, G, and B color channels for a digital image that belongs to the PH2 dataset. (d–i) An illustration of the tessellation patterns for each color channel. The triangle shapes of k = 1 are presented in (d–f); the fractal scaling parameter k = 4 is shown in (g,i). They are used to compute the $$X_{k_c}^m$$ matrices. (d,g) are R channel images. (e,h) are G channel images. (f,i) are B channel images.

Figure 5

The structure of an RBFNN classifier.

Figure 6

The prediction performance for the 5-fold cross validation and kNN classifier for different average percentage of color areas/color cluster descriptors.

Figure 6

The prediction performance for the 5-fold cross validation and kNN classifier for different average percentage of color areas/color cluster descriptors.