Technology Research Based On Color Asymmetry For Melanoma Detection

Melanoma is a kind of fast movement and great harmfulness malignant tumor，which moreoften occurs in the skin. In recent decades, the incidence and mortality of melanoma have a rapidgrowth trend. Therefore, early diagnosis of malignant melanoma are crucial for people to improvetheir life and healthy. In clinic, according to the surface properties and growth characteristics ofmelanoma, dermatologists summarize a useful principle for early detection - ABCD rule: TheA-rule is asymmetry, including boundary, color and texture; the B-rule is irregularity of contour; theC-rule is diversity of color; the D-rule is increasing of diameter.Computer aided diagnosis (CAD) technique of the melanoma can help doctor to distinguishbetween benign and malignant tumor and improve the accuracy rate of diagnosis. Based on thecolor asymmetry for clinical manifestations of skin melanomas, this paper provides technicalsupport for computer-aided diagnosis of skin cancer- malignant melanomas, using image processingand pattern recognition technology to research the analysis of color asymmetry, feature extractionand classification.In this paper, the main work and contributions are as follows:(1) Aiming to analyzing gray surface asymmetry of skin tumors, a modified virtual contourmodeling the surface asymmetry is proposed in the paper. In addition, a novel feature cluster,irregular and asymmetric descriptors of virtual contour, is constructed for measuring gray surfaceasymmetry of skin tumors. The experimental results show that the virtual contour on modifiedtransformation model can effectively enhance expression of tumor surface asymmetry.(2) Aiming to analyzing color asymmetry of skin tumors in the RGB space, the combination ofPrinciple Component Analysis (PCA) for dimension reduction and the modified virtual contourtransformation model is proposed. Firstly, the dimension of color images for skin tumors is reducedusing PCA transformation. Then, according to the modified virtual contour transformation modelproposed in the above, every principle component is converted to a virtual contour. Furthermore,the geometry asymmetry of virtual contour about color asymmetry axis is measured to reflect colorasymmetry of skin tumor surface.(3) Aiming to analyzing color asymmetry of skin tumors in the RGB space, color segmentationand wavelet decomposition is used to measure color asymmetry of tumor surface in PCA space. Anovel method of the unsupervised automatic color segmentation for lesion surfaces is presented inthe paper. And wavelet decomposition is suggested to measuring the asymmetry of each color. Theexperiment shows that color segmentation details color asymmetry analysis of skin tumor surface and covers the diversity of surface color. At the same time, compared with the comparative methodbetween pixels on mirror symmetry, the calculation is reduced and the asymmetry analysis on thecombination of approximate and detailed component is more accurate for the waveletdecomposition method.(4) In the classification of skin tumors, a small sample classifier is designed. A certain numberof benign and malignant samples are randomly selected from training set. In that case, we canobtain a variety of combinations from which fifteen kinds of combination are chosen. Two methodsbased on the average of classification results and the average of Receiver Operator Characteristiccurves of fifteen small sample classifiers are proposed to assess the performance of small sampleclassifier in this paper. The less number of tumor sample is resolved and the performance of theclassifier is improved effectively by the design of small sample classifier.