Development of a Dermatological Workstation: Preliminary Results on Lesion Segmentation in Cie L∗a∗b∗ Color Space

In the past we have proposed a pseudo-colorimetric imaging system for dermatological uses. The system exhibits moderate accuracy, but high precision or reproducibility. Acquired images are stored in the welldefined gamma-corrected sRGB color space. This color space has a known relationship to the perceptually uniform CIE L∗a∗b∗ color space, which allows the development of segmentation methods based on perceptual color differences. This should give segmentation results which are more in agreement with a human observer than if performed in uncalibrated RGB or other tristimulus color spaces. To demonstrate this a very simple semi-automatic segmentation method is proposed and tested, and it is hoped that more sophisticated methods, e.g. watersheds and clustering, may equally benefit from a similar colorimetric approach. The basic premise of the proposed segmentation method is that a skin lesion is surrounded by a band of ‘normal’ skin. The colors of this band are determined when the user draws a rough outer contour of the lesion, and are used to shrink the contour if the local color is similar to the local ‘normal’ skin color. As in general there is no consensus between dermatologists about the lesion contour, testing the method was primarily concerned with the reproducibility of the segmentation method versus that of a manually drawn contour. This reproducibility was assessed using two possibly diagnostically relevant parameters computed from the contours: the area in pixels, and the average L∗a∗b∗ color inside the contour. The test was conducted using a set of 30 distinct images, each used twice for a total of 60 segmented images per observer. Each segmented image was also scored by the observer. Four observers took part in the tests, and general satisfaction with the method was good, This work was supported by the Belgian FWO, ref. FW7801 although the presence of hairs still posed some problems. Reproducibility was not really improved compared to manual segmentation, as long as one and the same observer was concerned. However, with multiple observers working on the same images the results of the algorithm improved, mainly due to the poor reproducibility of manual contours between observers. Lastly, the method is very fast and needs only minor interaction to operate properly.