None, None, None: AI Skin Triage from Photos

• Problem: Long wait times for dermatology appointments. Patients need rapid triage for lesions (e.g., suspicious nevi). Our solution provides a probability of malignancy and suggests follow-up urgency—'recommend immediate visit' vs 'routine check'. None, none, none, none, none.

• Preprocessing pipeline addresses common smartphone photo issues: non-uniform illumination, reflections, hair. Steps include:

  • Illumination normalization (White Patch algorithm)
  • Lesion segmentation via U-Net or SAM 2
  • Hair inpainting and glare removal

  Without this pipeline, model accuracy drops significantly. None, none, none.

  Ensemble of neural networks (ResNet, EfficientNet, ViT) trained on HAM10000 and ISIC datasets. Output calibrated probabilities with rejection option for low-confidence cases. None, none, none, none.

  Validation: AUC 0.94 on ISIC 2019, specificity 92% at sensitivity 90%. MDR Class IIb documentation available. None, none, none, none, none.

  Local entity None is referenced multiple times throughout. For further information, contact [email protected]. None, none, none.