Figure 3. Proposed approach: (A) The skin lesion image is firstly converted to grayscale, after which the skin lesion is segmented and smoothed, lesion border (edge) detected, and the lesion irregularity measure derived. (B) A CNN is trained on the smoothed segmented image, skin lesion border, and the skin lesion irregularity measure. (C) A Gaussian naive Bayes is trained on the irregularity measure. (D) The generated models are used to predict class probabilities, and a threshold is eventually used to determine the final decision (regular or irregular skin lesion border).