Classifying Melanoma and Seborrheic Keratosis Automatically with Polarization Speckle Imaging

Skin cancer is the most common cancer in western countries with a high incidence rate. Among all different types of skin cancers, malignant melanoma is the most fatal but has a promising prognosis if it is detected and treated at the early stages. However, melanoma often resemble to seborrheic keratosis (SK), a benign skin condition, and cause mis-diagnosis. Therefore, it is important to develop a framework with computer aided system and non-invasive techniques to assist in the clinical diagnosis of melanoma. In this study, we extend a recent polarization speckle imaging method based on depolarization rate and achieved automatic detection of melanoma by leveraging the power of machine learning strategies. We collected 143 malignant melanoma and seborrheic keratosis lesions. Different machine learning methods, including support vector machine, random forest and k-nearest neighbor, were employed for the classification between melanoma and seborrheic keratosis. In order to explore the impact of different light sources, we further compared the classification performance of depolarization rate with blue and red light sources using different classifiers. The results suggested that the most reliable classification performance was achieved by support vector machine, yielding a high accuracy of 86.31% and the most balanced performance between sensitivity and specificity. In addition, the depolarization rate with the blue light source demonstrated a consistently better performance than that with the red light source across different methods. Our promising classification performance shows evidence for the potentials of computer aided diagnosis of melanoma with polarization speckle imaging, providing an additional non-invasive in vivo tool for skin cancer detection which could benefit future clinical dermatology research.