Determining the asymmetry of skin lesion with fuzzy borders

Abstract

It is highly desirable to identify malignant melanoma, a common cancer, at an early stage. One important clinical feature of this cancer is asymmetrical skin lesions. In this paper, we propose an adaptive fuzzy approach that uses symmetric distance (SD) to measure lesions with fuzzy borders. The use of a number of SD variations and the adoption of a backpropagation neural network enhances the discriminative power of the approach. Digitized images from the Lesion Clinic in Vancouver, Canada, demonstrate the accurate classification of asymmetric lesions at around 80%.

Introduction

Without timely diagnosis and treatment, malignant melanoma, a type of skin cancer, is frequently fatal [1], yet while early detection and surgical removal of melanoma greatly increases the chances of recovery, false-positive diagnoses increase health care costs and can lead to unnecessary surgery. Guidelines exist for the recognition of melanocytic lesions—such as the ABCD rule [2] and the Seven-point Checklist [3]—but even trained dermatologists find the diagnosis of lesions difficult. A clinical tool for the automatic diagnosis [4] of lesions would be very valuable.

There are three types of melanocytic lesions: benign nevi, dysplastic nevi and malignant melanomas. Benign nevi are usually less than $\text{5}\text{mm}$ in diameter, are a consistent tan-brown, and have a regular or distinct border. They do not develop into cancer. Dysplastic nevi are larger than $\text{5}\text{mm}$ in diameter, are variegated, ranging from tan to dark brown on a pink background, and their borders are usually irregular and frequently ill-defined. They present a higher risk of cancerous development, and are often characterized as precursors to cancer. The last type of lesion is malignant melanoma, characterized by sudden or continuous changes in size and the development of an irregular and/or notched border. These are cancerous.

The ABCD rule [5] is a set of diagnosis guidelines for the identification of malignant melanomas. Each letter in the rule represents a property: A stands for asymmetry, B for border irregularity, C for color variegation and D for diameter. Most malignant melanomas are black, blue, or red, asymmetric and irregular, and tend to grow. In a previous study, 12 different features deviated from the rule, such as asymmetry, color intensity and irregularity, were applied to a sample set of images of skin lesions. Two features were found to be significant, circularity and fractal dimension [6], [7], [8].

In our previous papers, we have shown that hairs on skin lesions can confuse the analysis of skin lesion's images, and we have successfully developed a program called DullRazor, which can perform hair removal with little destruction of the part of lesion in images [9]. Also, a multi-median filter has been proposed and used to identify the location of skin lesions from images [10].

This paper continues our previous work investigating how to identify asymmetry in the ABCD rule. In order to verify the efficiency of the asymmetric measurement of skin lesions, we have done the following:

• Studied and compared different asymmetric measurements to determine how these measurements are effective to measure the asymmetry of skin lesions with fuzzy borders.

• Analyzed how point selection and definition of symmetry axis affect the accuracy of the measurements.

• Proposed a method to handle border fuzziness and enhance the discriminative powers of the measurements.

• Derived a backpropagation neural network to improve the discriminative powers of the measurement results.

Different approaches have been proposed to make the identification and clinical diagnosis of malignant melanomas more accurate. In order to produce useful lesion features, Lizzi et al. [11] used therapeutic ultrasound transducers to detect the thermal condition of lesions. To analyze the spatial distribution of intensity to differentiate between melanomas and other melanomas lesions, Chwirot et al. [12] used digital images of auto-fluorescence, although, while there are different wavelengths for different methods, they mainly analyzed the heating of tissues with electromagnetic devices. A common problem for this kind of method is that improper application of heat may destroy skin tissues.

On digitized images of skin lesions, Lee et al. [13], [14] used a new measurement called irregularity index to measure their border irregularity. This index locates all the local and global indentations and protrusions and organizes them in a hierarchical structure. The irregularity index is computed for each indentation and protrusion along the border, and the most significant and overall index are derived and used to estimate the irregularity. This method does not, however, take into account the border's fuzziness. Denton et al. [15] used the progressive filter to locate a lesion's border. They progressively increased the strength of the filter to exclude pixels with low color intensities, then, when the differences between two successive filtered lesions are less than a pre-set threshold, the skin lesion can be identified. However, this method is sensitive to the scale of lesions since the change of color intensity at the border is sharp. Another disadvantage is that this method detects the borders by eliminating the pixels at lesion's borders, so the detected lesions may in fact be significantly different from the original shapes.

Stoecker et al. [16] studied the identification of the asymmetry of malignant melanomas using the symmetric distance (SD). However, while this is based on the exact identification of the lesion's border, it does not include an SD measure of border fuzziness, rendering the calculated SD sensitive to the segmentation results. Other studies have measured infrared and ultraviolet reflectance [17], used a low-power surface microscope [18], digitized and analyzed the pre-existing slides [6], and constructed a spectroanalyzer [19].