Background: The incidence rates of melanoma have risen to worrying levels over the last decade. Delayed diagnosis, due to faults on the detection stage, indicates the necessity of new aiding diagnosis techniques. Since metabolic activity is highly connected to neoplasia formation, a detection technique that focuses its results on vascular responses, as Infrared thermal (IRT), seems to be a viable option.
Materials and methods: Static and dynamic (cooling) thermal images of melanoma and melanocytic nevi lesions were collected and analysed to retrieve thermal parameters characteristic of this skin lesion types. The steady-state and dynamic variables were tested separately with different machine learning classifiers to verify whether the distinction of melanoma and nevi lesions was achievable.
Results: The differentiation of both types of skin tumours was doable, achieving an accuracy of 84.2% and a sensitivity of 91.3% with the implementation of a learner based on support vector machines and an input vector composed by static variables.
Conclusion: The use of IRT for skin tumour classification is achievable, but some improvement is needed to raise the metrics of sensitivity and specificity. For future work, it is recommended the study of dynamic parameters for the classification of other types of skin neoplasia.
© 2019 European Academy of Dermatology and Venereology.