Development of convolutional neural networks for skin disease diagnosis with optimization methods

In recent years, skin cancers have become one of the most common and deadly diseases worldwide. Early diagnosis plays a critical role in preventing disease progression and improving treatment success. However, the visual similarity of different skin lesions makes the diagnostic process challenging for dermatologists. Therefore, computer-based image analysis and deep learning–based approaches have become an important research area for the automatic diagnosis of skin diseases. In this study, a deep learning model was developed for the classification of skin diseases using the ISIC 2019 dataset. In the first stage, five different pre-trained architectures were evaluated through transfer learning, and EfficientNetV2S was selected as the base model for achieving the highest F1 score. In the second stage, the EfficientNetV2S model was enhanced with channel and spatial attention mechanisms, improving its ability to select meaningful features and emphasize important regions within the images. In the third stage, the hyperparameters of the enhanced model were optimized using the Equilibrium Optimizer (EO) algorithm. The EO algorithm effectively improved the model's performance, demonstrating its capability as a powerful approach for hyperparameter optimization. Finally, the model was retrained on a larger dataset, which led to a further increase in the F1 score. The results indicate that the integration of attention mechanisms, EO-based hyperparameter optimization, and training with a larger dataset significantly improve the performance of deep learning models for skin disease classification.