TRANSILLUMINATION IMAGING FOR EARLY SKIN CANCER DETECTION

[attachment=17586]

I. INTRODUCTION
Frequent screening of suspicious skin pigmentations is a very effective approach for detecting skin cancers before they become lethal, since early changes in a malignant nevus typically consist in the development of an irregular pigmentation pattern. However, visual detection of melanomas by a dermatologist has an average diagnostic accuracy of only 58% [4,7], but it can be improved using imaging techniques [31,40], such as oil-immersion [4] and cross-polarization [5] epiluminescence imaging (ELM). In general, these imaging techniques rely on delineating the boundary of a lesion, which is then analyzed for certain skin pigmentation characteristics, such as shape, size, symmetry, color, and texture. To this effect, a number of image segmentation techniques have been proposed in the past several years with varying degrees of success.


II. METHODS

A. Data Description
All of the lesions analyzed in this study were imaged using a Nevoscope device capable of obtaining both TLM and XLM images. The device used an optical lens (Nikon, Japan) to achieve a standard 5X magnification and an Olympus C2500 (Olympus, Japan) digital camera for capturing the images. For each lesion two distinct images were acquired, one in the TLM and one in the XLM modality. Clinical imaging was carried out at the University of Texas in Houston under the direction of a board certified physician (MD). Full institutional approvals were obtained for all studies.
A total of 60 lesions were imaged from consecutively and prospectively enrolled clinic patients undergoing routine skin exams, who had clinically suspicious skin lesions of less than 1 cm in size. To avoid clinician bias, all dermatology patients were considered for participation in the study regardless of risk of melanoma or past history.

B. Procedure Outline
The ELM and XLM images undergo completely automated analysis that consists of five main steps, before the lesion in the images can be extracted, classified, and displayed with the lesion boundary outlined. In turn, each main step consists of several sub-steps. The entire procedure is graphically depicted in Figure 1. Initially, the images are preprocessed and segmented with three different methods out of which the final lesion area is selected by a scoring stage. Then, the lesion is classified as malignant or benign by comparing the areas of the lesion in the TML and XTL images. Finally, the refined boundary of the lesion is plotted on the original images.