Calibration-Aware Heterogeneous Multimodal Ensemble for Multi-Class Skin Lesion Diagnosis

Background: Reliable classification of skin lesions is essential for early melanoma detection; however, substantial visual overlap among lesion types and pronounced class imbalance continue to limit the performance and reliability of existing automated systems. These challenges motivate the integration of complementary visual representations with clinical metadata to improve discrimination across lesion classes. Methods: This study introduces a calibrated multimodal ensemble learning framework for skin lesion classification. The proposed framework combines visual features extracted from Swin Transformer, EVA-02, and CoAtNet architectures with patient age and anatomical region information. A two-stage adaptive temperature scaling is applied to calibrate the predicted probabilities of the ensemble. Results: Experimental evaluation on the International Skin Imaging Collaboration (ISIC 2019) dataset demonstrates strong classification performance, achieving a macro-F1 score of 0.91 and a macro-AUC of 0.99. The framework further showed robust generalization on the official ISIC 2019 benchmark set, attaining a balanced accuracy of 0.65. The external validation on the PH2 dataset yielded a macro-F1 score of 0.93 and an ROC-AUC of 0.96. Conclusions: These findings demonstrate that combining heterogeneous visual representations, clinical metadata, and calibration-aware ensemble learning provides a robust and trustworthy solution for skin lesion classification and computer-aided dermatological diagnosis.