An online cutaneous melanoma risk screen tool.

To develop and deploy a publicly accessible online risk estimation tool for cutaneous melanoma that prioritizes high recall to minimize missed diagnoses, using real-world clinical photographs rather than dermoscopic images, and to bridge the gap between artificial intelligence research and clinical implementation. A 'segmentation-first, then-classification' strategy was adopted. Lesions were segmented using MFSNet pretrained on dermoscopic images. The classification model was customized from EfficientFormerV2-L by integrating dual attention, multi-scale pooling, and feature fusion modules. A balanced dataset of 934 clinical images (467 melanoma, 467 nevus) from PAD-UFES-20, Seven-Point Check, Med-node, and other sources was used. Weighted BCEWithLogitsLoss (positive weight up to 6) was applied to increase the penalty for false negatives. External validation was performed on 32 smartphone-captured images (16 melanoma, 16 nevus). The model was deployed online with an automated image quality check. The customized model (trained from scratch, positive weight 6) achieved a recall of 0.95, precision of 0.75, and F1 score of 0.84 on cross-validation, outperforming the base EfficientFormerV2-L (recall 0.88). Dermoscopic pretraining unexpectedly degraded classification performance on clinical images. External validation showed significantly higher malignant potential scores for melanoma versus nevus cases (P < 0.0001). The online tool is freely available at http://www.make-a-difference.top/mmscreen. This study provides a practical, high-recall online melanoma screening tool for real-world clinical images. Its free public access and emphasis on minimizing false negatives address critical gaps in early melanoma detection, particularly in primary care settings.