Clinically Reliable Melanoma Classification from Darmoscopy Image Using Linear and Nonlinear MIL Approaches

Melanoma is one of the most aggressive forms of skin cancer, accounting for an excessive number of skin cancer-related deaths despite its relatively low incidence. Early and accurate detection is crucial to improving survival rates, but manual diagnosis through dermoscopy remains challenging due to the subtle lesion characteristics. In this study, we propose a lightweight and resource-efficient framework for melanoma detection that integrates image segmentation with multiple instance learning (MIL)-based classification. Dermoscopic images from the ISIC 2018 dataset were preprocessed using resizing, noise removal, and hair artifact reduction to improve visual consistency. Lesion segmentation was performed using a statistical measure of pixel intensity standard deviation (SD), which adaptively determined the most suitable enhancement and thresholding techniques across three criteria. Segmentation outputs were then used to extract region-specific features, which were classified using three MIL-based approaches: mi-KSPSVM, mi-SPSVM, and MIL with Lagrangian relaxation, along with standard SVM baselines. Model performance was evaluated under 5-fold, 10-fold, and Leave-One-Out (LOO) cross-validation schemes using accuracy, sensitivity, specificity, F1-score, and computational efficiency as metrics. Experimental results demonstrate that segmentation significantly enhances classification performance, particularly in terms of sensitivity and F1-score. Notably, mi-KSPSVM achieved the highest sensitivity of 94.74% under 10-fold CV, while mi-SPSVM exhibited superior generalization under LOO validation. Additionally, preprocessing and segmentation consistently reduced misclassification of melanoma cases while preserving computational efficiency, with the SVM (RBF) baseline remaining the fastest model. Future work will explore the integration of deep learning-based segmentation to improve accuracy further and extend applicability to other skin cancer subtypes.