Melanoma Diagnosis with Multiple Decision Trees

This chapter highlights the application of multiple binary decision trees in melanoma diagnosis. Since the clinical rules in diagnosing melanoma involve inhomogeneous/non-metric data and various ‘if-then’ statements, direct utilization of machine learning techniques such as neural networks can not perform satisfactorily in modelling the clinical diagnostic knowledge which, typically, is nonlinear and fuzzy. As a versatile and intuitive paradigm in pattern classification, the decision tree is perhaps the optimal mechanism in mimicking the clinical diagnostic rules. This chapter compares the performances of two different designs of the multiple decision trees via experiments. Digital image attributes, including both geometric and colorimetric ones, are all examined in detail. Receiver operating characteristic curves of varying ensemble sizes are presented, illustrating the effectiveness of decision trees in melanoma diagnosis.