Improving Skin Lesion Classification Accuracy Using VGG19 Combined with Optimized Random Forest of Uncorrelated High-Performing Trees

Abstract

Accurate and early prediction of skin cancer is essential for improving survival rates and minimizing the burden of treatment, as timely diagnosis greatly enhances patient outcomes. Automated classification methods have become invaluable tools for dermatologists, offering consistent, rapid, and objective analysis of skin lesions. Given the prevalence and potentially fatal nature of skin cancer, developing accurate and efficient classification systems is crucial for effective early detection and treatment. This study proposes an optimized Random Forest (RF) framework combined with VGG19-based feature extraction to enhance classification performance and computational efficiency. Features extracted from VGG19 were used to train N decision trees, and the top P high-performing trees were selected based on their accuracy. To ensure diversity and reduce redundancy, Jaccard distances were computed, and Q uncorrelated trees were identified using K-Means clustering. These optimized trees formed an ensemble, and predictions were aggregated through majority voting. The proposed framework achieved 88.21% accuracy, 81% precision, 93% recall, and 87% F1-score, significantly outperforming other state-of-the-art models.