A Novel Hybrid Watershed and Extreme Learning Machine Framework for Skin Cancer Classification

This study presents an efficient framework for skin cancer segmentation using Watershed algorithm and classification using Extreme Learning Machine model (ELM) with Histogram of Oriented Gradients (HOG) feature extraction and Principal Component Analysis (PCA) for dimensionality reduction. Segmentation stage exhibits a strong performance indicated through good Dice coefficient and precision value. The classification algorithm achieves 93.5% test accuracy with 91.3% sensitivity and 95.0% specificity on a melanoma classification dataset, demonstrating strong diagnostic capability while maintaining computational efficiency. The PCA reduction preserves 95% variance, enabling the lightweight ELM architecture to train 23 times faster than conventional deep learning approaches while maintaining competitive performance as given by the F1-score of 0.92. Brier score of 0.16 indicates a well calibrated probabilistic output while high negative predictive value suggests reliable prediction. These results suggest that the ELM-PCA-HOG combination offers an effective balance between accuracy and efficiency for clinical decision support systems, particularly in resource-constrained settings.