Skin Lesions Asymmetry Estimation Using Artificial Neural Networks

Artificial Neural Networks (ANNs) are efficient tools successfully used to solve a regression problem. In this paper, the skin lesions are analyzed using a feedforward neural network (FFN) with Levenberg-Marquardt Backpropagation (LMBP) training algorithm as a supervised learning method. The proposed model uses four combinations of inputs built on the data from type of skin lesion/database/ method of asymmetry computation and searches for four combination of desired outputs such as the type of skin lesion/database/ method of asymmetry computation. Also, the number of hidden neurons has been changed to reach the condition of maximum regression coefficient (R) and minimum mean squared error (MSE). The proposed FFN-LMBP model was validated with 24 images and tested with another 24 images. This study is centered on the most relevant and widely used feature in dermoscopic images, i.e., asymmetry. Two algorithms are implemented to extract handcraft asymmetry values: one algorithm computes the asymmetry of the geometric characteristics (GAF) using the geometric shape of the lesions, and the second one computes the asymmetry based on histogram projections (AHP) on the horizontal and vertical axes. The MED-NODE and PH2 databases are used for skin lesion detection.