Data Processing Beyond Visual Interpretation

Electronic classifier systems (machine learning models) are increasingly being trained for the diagnosis of pigmented skin lesions, using any number of imaging modalities to provide training data. The performance of such systems in laboratory settings is at about the same level as human experts. This paper gives some background information on modern classifier systems, distinguishing between white-box and black-box systems. Furthermore, it stresses the need to combine lesion-based information with prior information about the health state of the patient to arrive at a more accurate diagnosis. To this end, we propose to use Bayesian networks to model the prior probability of disease in a patient, and Bayesian reasoning to combine this prior information with the output of machine learning models.