Skin tumor identification by means of convolutional neural network and improved gray wolf optimizer.

Introduction: Early and accurate diagnosis of skin cancer has a huge impact on the survival rate of patients and deep learning-based CNN and dermatologists' intelligence support can fasten the diagnosis not only within the clinics but also outside. Methods: This paper introduces a hybrid classification framework for automatic detection of melanoma that combines deep learning and better optimization. A modified GWO method called improved GWO (IGWO) has been proposed which can efficiently optimize CNN parameters and feature learning. Conventional GWO has drawbacks such as early convergence and poor exploration in high-dimensional spaces, so the IGWO regenerates the weak omega agents based on their fitness level. The underperforming omega wolves were discarded in every step and then either the elite solutions (alpha, beta, delta) or the stochastically resampled solution was put to the wolves so that both exploration and exploitation would reach better state. Results: The presented CNN/IGWO model has been experimented with a skin cancer dataset SIIM-ISIC 2020. The proposed model yielded test accuracy of 98.47% and AUC of 98.2 respectively that were both higher than those models using basic GWO method and other state-of-the-art deep learning methods. Discussion: These outcomes show that using the IGWO mechanism along with training of CNN speeds up convergence, enhances the solution and thus the classification performance. The proposed method shows that intelligence-based optimization could give more practical, accurate results in automated melanoma diagnosis.