Feature-based Differentiation of Malignant Melanomas, Lesions and Healthy Skin in Multiphoton Tomography Skin Images

Abstract

Malignant melanoma is a very aggressive tumour with the ability to metastasize at an early stage. Therefore, early detection is of great importance. Multiphoton tomography is a new non-invasive examination method in the clinical diagnosis of skin alterations that can be used for such early diagnosis. In this paper, a method for automated evaluation of multiphoton images of the skin is presented. The following features at the cellular and subcellular level were extracted to differentiate between malignant melanomas, lesions, and healthy skin: cell symmetry, cell distance, cell density, cell and nucleus contrast, nucleus cell ratio, and homogeneity of cytoplasm. The extracted features formed the basis for the subsequent classification. Two feature sets were used. The first feature set included all the above-mentioned features, while the second feature set included the significantly different features between the three classes resulting from a multivariate analysis of variance. The classification was performed by a Support Vector Machine, the k-Nearest Neighbour algorithm, and Ensemble Learning. The best classification results were obtained with the Support Vector Machine using the first feature set with an accuracy of 52 % and 79.6 % for malignant melanoma and healthy skin, respectively. Despite the small number of subjects investigated our results indicate that the proposed automatic method can differentiate malignant melanoma, lesions, and healthy skin. For future clinical application, an extended study with more multiphoton images is needed.

Keywords

• skin neoplasm
• nevus
• skin cancer diagnosis
• multiphoton fluorescence microscopy
• biomedical image processing
• machine learning