Skin cancer, especially melanoma, has become one of the most widespread and deadly diseases today. The chances of successful treatment are greatly reduced if the melanoma is not treated in its early stages because it could spread aggressively. Hence, the diagnosis of skin cancer is very challenging as skin lesions are highly subjective to analyze and that type of expertise is exceedingly specialized. While there is an increase in the prevalence of skin cancer across the globe, there is an increase need of automated diagnostic systems that could aid medical personnel in making appropriate decisions within the requisite timelines. This study proposes construction of a deep learning model built on the MobileNetV2 architecture that has been memetic optimized for hyperparameter tuning. The memetic algorithm employs both global and localized search techniques to fine-tune the model parameters that include learning rate, batch size, and number of epochs to boost the efficacy of the model. This makes it possible for the proposed model to achieve high performance while remaining economical on resources. This makes the model suitable for real world clinical settings. The model achieved exceptional results, with 98.48% accuracy, 97.67% precision, and 100% recall, highlighting its strong ability to detect malignant lesions. The ROC AUC score of 99.79% further demonstrates its outstanding capability to differentiate between benign and malignant lesions. Notably, visualizations such as the Grad-CAM heatmap and Superimposed Image were crucial in providing interpretability to the model's decision-making process. The Grad-CAM heatmap highlighted the regions of interest in the lesions, showing how the model focused on key structural features. The Superimposed Image combined these heatmaps with the original lesion images, making it clear which parts of the lesions influenced the model's classification. These results underscore the potential of deep learning models, optimized with the memetic algorithm, to significantly improve skin cancer detection. By offering both high accuracy and interpretability, this model presents a valuable tool for dermatologists, facilitating faster and more reliable early diagnosis and ultimately improving patient outcomes.
Şençelikel et al. (Thu,) studied this question.