A SVM-based diagnosis of melanoma using only useful image features

In this study, we propose an automated system for detecting melanoma skin cancer from plain photographs of affected skin regions. Melanoma is the deadliest of skin cancer types and cases of its occurrence continues to rise. Like most cancers early detection is vital in improving the chances of survival. Computer aided diagnoses using digital image processing can assist the skin doctor in identifying melanoma because it occurs mainly on the body exterior. In most cases ABCDEs rule has been applied for detecting melanoma, and therefore we apply similar method. We first segment an input image into lesions of interest appeared to be melanoma by GrabCut algorithm, and next extract some features such as the shape, color, and geometry by using image processing techniques. These extracted features are categorized as cancerous "malignant" or non-cancerous mole "benign" by using support vector machine with Gaussian radial basis kernel (SVM-RBF). We conducted evaluation experiments with 200 images (100 of melanoma and 100 of benign) and found from the results that only six features can be sufficient to detect melanoma.