Cascade Classification of Dermoscopic Images of Skin Neoplasms with Controllable Sensitivity and External Clinical Validation

The study compares four deep learning architectures—ViT-B/16, Swin-S, ConvNeXt-S, and EfficientNetV2-S—on dermoscopic images for skin neoplasm classification using binary, single-stage, and a two-stage cascade approach. The cascade model improves macro F1 scores by enhancing sensitivity, particularly for malignant lesions, but reveals a generalization gap when transferring from open datasets to clinical settings, necessitating external validation and recalibration. This research highlights the importance of architecture selection and classification strategies in achieving reliable performance in real-world clinical applications.