The need for diagnosing malignant melanoma in its earliest stages results in an increasing number of unnecessary excisions. Objective criteria beyond the visual inspection are needed to distinguish between benign and malignant melanocytic tumors in vivo. Fluorescence spectra collected during the prospective, multicenter observational study ("FLIMMA") were retrospectively analyzed by the newly developed machine learning algorithm. The formalin-fixed paraffin-embedded (FFPE) tissue samples of 214 pigmented skin lesions (PSLs) from 144 patients were examined by two independent pathologists in addition to the first diagnosis from the FLIMMA study, resulting in three histopathological results per sample. The support vector machine classifier was trained on 17,918 fluorescence spectra from 49 lesions labeled as malignant (1) and benign (0) by three histopathologists. A scoring system that scales linearly with the number of the "malignant spectra" was designed to classify the lesion as malignant melanoma (score > 28) or non-melanoma (score ≤ 28). Finally, the scoring algorithm was validated on 165 lesions to ensure model prediction power and to estimate the diagnostic accuracy of dermatofluoroscopy in melanoma detection. The scoring algorithm revealed a sensitivity of 91.7% and a specificity of 83.0% in diagnosing malignant melanoma. Using additionally the image segmentation for normalization of lesions' region of interest, a further improvement of sensitivity of 95.8% was achieved, with a corresponding specificity of 80.9%.
Keywords: dermatofluoroscopy; machine learning; malignant melanoma; melanin fluorescence; support vector machine.