SVM-based Texture Classification and Application to Early Melanoma Detection

We have recently developed a decision support system for early skin cancer detection that relies on analysis of the pigmentation characteristics of a skin lesion, detected using crosspolarization imaging, and the increased vasculature associated with malignant lesions that is detected using transillumination imaging. Current system uses size difference based on lesion physiology and achieves great overall accuracy (86.9%). In this paper, we explore texture information, one of the criteria dermatologists use in the diagnosis of skin cancer, but has been found very difficult to utilize in an automatic manner. The overarching goal is to improve the overall decision support capability of the DSS. The objective is to use texture information ONLY to classify the benign and malignancy of the skin lesion. A three-layer mechanism that inherent to the support vector machine (SVM) methodology is employed to improve the generalization error rate and the computational efficiency. The performance of the algorithm is validated with a series of benchmark texture images and then tested on 22 pairs of real clinical skin lesion images. Our experimental results show that a 4^th-order polynomial kernel can reach an average accuracy of 70% in determining the malignancy of any pixel within any given skin lesion image. Further study will look at whether multi-channel filtering based feature extraction algorithm will improve the accuracy rate, and the performance comparison between SVM-based texture classification and decision tree-based texture classification in both the spatial and frequency domain