About FGH-Dermatology
FGH-Dermatology exists to close a well-documented equity gap in dermatological AI: most diagnostic models underperform on darker skin tones. This platform is built to support clinicians assessing patients across the full Fitzpatrick spectrum, with every prediction accompanied by visual and natural-language explanations so the reasoning is always auditable.
Model Architecture
The platform runs a convolutional neural network supplied as a trained Keras .h5 file, loaded once at API startup. The network outputs a softmax probability distribution across four diagnostic classes.
| Input resolution | Auto-detected (default 224×224) |
| Output classes | 4 diagnostic categories |
| Format | HDF5 (.h5) |
| Explainability | Grad-CAM + NLP reporting |
| Model hosting | Hugging Face Hub |
Supported Diagnostic Classes
| Code | Class | Risk |
|---|---|---|
| MEL | Melanoma | HIGH |
| BCC | Basal Cell Carcinoma | HIGH |
| NV | Melanocytic Nevus | LOW |
| BKL | Benign Keratosis-like Lesion | LOW |
Explainability Methods
Grad-CAM highlights the spatial regions of an input image that most influenced the model's prediction, letting a clinician verify whether the model attended to the lesion rather than background.
Natural language reporting translates the prediction, confidence level, Grad-CAM activation regions, and skintone context into a structured clinical summary including differential considerations and recommended next steps.
Intended Use & Contraindications
FGH-Dermatology is intended for use by qualified clinicians as a decision support aid alongside — never in place of — full clinical history, physical examination, and dermoscopic assessment. It is not validated for autonomous diagnostic use, patient self-assessment, or use as the sole basis for a treatment decision.
References
- Selvaraju, R. R., et al. "Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization." ICCV, 2017.
- Tschandl, P., Rosendahl, C., & Kittler, H. "The HAM10000 dataset." Scientific Data, 2018.
- International Skin Imaging Collaboration (ISIC) Archive — isic-archive.com
- Groh, M., et al. "Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fitzpatrick 17k Dataset." CVPR Workshops, 2021.