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.