Unsupervised Numerical Characterization in Determining the Borders of Malignant Skin Tumors from Spectral Imagery

For accurate removal of malignant skin tumors, it is crucial to assure the complete removal of the lesions. In the case of certain ill-defined tumors, it is clinically challenging to see the true borders of the tumor. In this paper, we introduce several computationally efficient approaches based on spectral imaging to guide clinicians in delineating tumor borders. First, we present algorithms that can be used effectively with simulated skin reflectance data. By using simulated data, we gain detailed information about the sensitivity of the different approaches and how variables defined by algorithms act in the skin model. Second, we demonstrate the performance of the algorithms with spectral images taken in-vivo and representing two types of skin cancers with ill-defined borders, namely lentigo maligna and aggressive basal cell carcinoma. The results can be used as a guideline for developing software for the fast delineation of skin cancers.