R. Pompl, W. Bunk, and G.E. Morfill
(in cooperation with TU München (IMSE), FH München, Dermat. Klinik d. Univ. Regensburg)
Image analysis techniques developed for astrophysical questions (source detection, characterization of extended sources) are applied on microscopic skin images to improve early recognition of skin cancer.
The malignant melanoma is the most dangerous human skin disease. This form of cancer leads quickly to metastases. Its incidence has almost doubled within the last decade. At an advanced stage the prognosis is very bad. On the other hand, if recognized early, the malignant melanoma can be excised and the patient can recover completely. These facts emphasize the importance of early recognition. In many cases it is hard to discriminate it from benign melanocitic nevi. One way to improve diagnostic accuracy at an early stage is dermatoscopy (skin surface microscopy at 10-times magnification), which yields skin images of the lesion that are rich of details. In order to guide the dermatologist in the difficult assessement of what he sees on these images, the semi-quantitative ABCD-rule has been developed. This rule evaluates four features of the lesion: Asymmetry, Border, Color and Different structural components. However, the successful application of this rule requires a great deal of experience. It is the purpose of this project to derive reproducible and quantitative measures for each diagnostic criterion, in order to quantify the classification of nevi with respect to their dignity. The image analysis includes also the pre-processing, such as segmentation of the lesion from the surrounding healthy skin and the removal of artefacts (e.g. hairs) in the digital image.
In addition to the classical image analysis methods (e.g. various filters and color space transforms), a number of new methods are used, which were developed at the MPE for extracting information from time series or from images. A kernel of the digital image analysis is the so-called Scaling Index Method (SIM) and its modification, the Scaling Vector Method (SVM), which are particularly suited to detect and quantify structural components or patterns in images. They are used for the removal of artefacts and the evaluation of the B and D component of the dermatoscopic rule. The examples show the application of the SIM to quantify the border region of a benign and malign lesion (melanoma). The benign example (left figure) is characterized by a thin and very regular shape. On the other hand, the width of the transition zone of the malign specimen (right figure) is irregular and the contour is rugged. These characteristics are quantified with aid of SIM: attached to each figure of the border region, the distribution of derived scaling-indices (p(alpha)-spectrum) is shown. The spectrum of the malignant melanoma has significantly different characteristics compared with that of the benign lesion. The results could be generalized in a pilot study of more than 110 records of suspected skin lesions. Up to now many classification-relevant features have been investigated and various measures for a valuation have been defined. Combination of all these measures of the nevi yields a recognition rate which is very close to that of an experienced dermatologist.