Evolutionary optimization for adaptive clustering of dermoscopy melanoma image

Automatic segmentation is one of the key steps for automatic analysis of melanoma images. An adaptive clustering algorithm, based on a combination of self-generating neural network (SGNN) with improved genetic algorithms (IGAs), is proposed for dermoscopy melanoma image in this paper. This algorithm involves three major steps: firstly, a group of optimal seed samples are selected through IGAs; and then these seed samples are taken as initial trees to generate the self-generating neural forest (SGNF) by training the rest samples based on SGNN; finally, every tree in the SGNF may denote a cluster in skin melanoma image to complete the clustering segmentation. In order to improve the performance of conventional GAs, the initial population size is firstly set according to the size of solution space, and then the genetic control parameters, such as population size, crossover probability and mutation probability, are adjusted during the evolving process, thereby the computational time is greatly shortened. Moreover, SGNN is combined with IGAs to overcome the sensitivity of SGNN to the trained order of samples so that the number of clusters can be determined adaptively without any manual intervention. The experimental results show the better stability and the satisfied clustering performance of the algorithm proposed in this paper.