Exploration of Genetic Algorithms and CNN for Melanoma Classification

Abstract

This contribution reports the first results we achieved by combining the main capabilities of Genetic Algorithms (GA) with Convolutional Neural Network (CNN) to address the melanoma detection problem (GACNN). The presented results are related to the melanoma classification problem we chose due to the large proposal available in literature usable for performance comparison. We used a clinical dataset (i.e., MED-NODE) as the data source. We compared performance obtained by GACNN with the AlexNet both with and without the Otsu segmentation. In addition, we considered the accuracy parameter as the scoring function for the GA evolution process. The preliminary results suggest the proposed approach could improve melanoma classification by allowing network design to evolve without user interaction. Furthermore, the suggested approach can be extended to additional melanoma datasets (e.g., clinical, dermoscopic, or histological) in the future with other innovative evolutionary optimization algorithms.