In this paper, we first propose a color image segmentation method based on the Dempster-Shafer's theory. The tristimuli R, G and B are considered as three independent information sources which can be very limited or weak. The basic idea consists in modeling the color information in order to have the features of each region in the image. This model, obtained on training sets extracted from the intensity, allows to reduce the classification errors concerning each pixel of the image. The proposed segmentation algorithm has been applied to biomedical images in order to detect a kind of skin cancer (melanoma). In a second step, features concerning the lesion are extracted using color information. These features are used in order to classify the beginning lesions (naevus) from the other. Results, including the management of false alarms and no detection, allow to demonstrate the effectiveness of the proposed methodology.