High-level features for automatic skin lesions neural network based classification

Melanoma is the most dangerous form of skin cancer. It can be developed from pigmented cells of the skin and can grow and spread swiftly to other organs (metastasis). An early diagnosis increases the chance of cure. In the past three decades, the increase in the incidence of melanoma has given rise to more accurate methods of analysis. Feature extraction is a critical step in melanoma decision support systems. Early dermatoscopic rules (ABCD rule, 7-point checklist, Menzies method and CASH algorithm), used by experts are generally low level features. In this paper, we consider several dermatoscopic rules for automatic detection of melanoma in order to generate new high level features allowing semantic analysis. Such extracted features are based on shape characterization and color and texture features. A neural network classifier is used for decision making. Experimental results indicate that semantic analysis is a useful method for discrimination of melanocytic skin tumors with good accuracy. The proposed method yields a good sensitivity of 92% and a specificity of 95% on a database of 206 skin lesion images. A comparative study with recent previous works illustrates that our approach outperforms in terms of accuracy and specificity.