Integrating features of radiomics and CNN models for early skin cancer detection based on watershed segmentation

Skin cancer is among the most aggressive and prevalent forms of cancer worldwide, with melanoma posing a high risk of metastasis and mortality when not detected early. Manual diagnosis by dermatologists, while effective, faces challenges such as subjectivity, variability, and limited accessibility in underserved regions. To address these limitations, this study proposes a robust computer-aided diagnostic system for early detection of skin cancer using a hybrid feature extraction approach and machine learning. In this study, several methodologies were developed for the automated classification of dermoscopic images, with a primary focus on a hybrid diagnostic model combining radiomic and deep learning features. Specifically, a Random Forest (RF) classifier was trained on fused features extracted from radiomic algorithms and deep convolutional layers of CNN. The proposed MobileNetV2 + radiomics + RF model achieved outstanding performance, with an average AUC of 85.09%, accuracy of 94.7%, sensitivity of 84.53% and specificity exceeding 99.2%. It exhibited particularly strong classification capabilities for high-risk lesions such as melanoma (AUC of 94.4%) and nevi (AUC of 98.2%), while maintaining robust performance across other lesion types. The integration of radiomic and CNN-based features through an RF classifier offers a highly effective approach for early skin cancer detection, with significant implications for clinical practice.