A fusion framework for melanoma detection using SkCanNet model and hand-crafted features

Skin cancer remains a major global health concern, emphasising the need for early and accurate diagnosis. Computer-aided diagnosis (CAD) has been applied for skin cancer detection. A critical issue in skin cancer detection is the detection and removal of artefacts like hair and noise from lesion images. Existing hair removal techniques often fail to preserve lesion details, affecting classification accuracy. To enhance melanoma detection, an advanced pre-processing method is necessary while preserving lesion features with integration of deep learning with handcrafted features. This paper presents an advanced pre-processing technique using Modified E-shaver, Modified Dull Razor, and Adaptive Principal Curvature. SkCanNet, is proposed for feature extraction which is combined with standard classifiers like KNN, SVM, and logistic regression for classification. The proposed approach significantly enhances skin cancer detection accuracy, specificity, sensitivity, and F1-score. Compared to existing deep learning models proposed system presents a promising solution for early melanoma detection.