A New Framework for Automated Skin Cancer Classi-fication based on Multifractal Feature Extraction us-ing Random Forest

Abstract

Skin cancer is a major public health concern, and early accurate classification is essential for improving diagnosis and treatment outcomes. However, differentiating basal cell carcinoma (BCC), squamous cell carcinoma (SCC), and melanoma remains challenging because of their visual similarity, irregular lesion borders, and heterogeneous morphological patterns. This study proposes an automated skin cancer classification framework based on multifractal feature extraction and Random Forest classification. A total of 1,043 dermoscopic images collected from ISIC and Kaggle repositories were used, including 454 melanoma, 392 BCC, and 197 SCC cases. The images were preprocessed through grayscale conversion, adaptive binarization, and region-of-interest preparation. Multifractal descriptors, including generalized dimensions and singularity spectrum features, were then extracted to quantify lesion complexity, spatial irregularity, and textural heterogeneity. These features were used to train a Random Forest classifier for three-class classification. The proposed framework achieved 98.8% accuracy, 98.8% sensitivity, 98.8% precision, 98.8% F-score, and 99.4% specificity. The results demonstrate that multifractal analysis combined with machine learning provides an accurate, interpretable, and computationally efficient approach for automated skin cancer classification..