Skin lesion diagnosis support, from a single image.
DermaScan AI is an artificial-intelligence system that diagnoses skin diseases from images, built to support clinics and pharmacies with fast, early detection of skin cancer and other lesions — and to send the right patients to a dermatologist sooner.
Skin disease is everywhere. Specialists are not.
Most skin complaints are first seen by a pharmacist, nurse or general practitioner — not a dermatologist. Telling a harmless mole from an early skin cancer by eye is hard, and the cost of a miss is high.
people are affected by a skin condition at any point in time.
leading cause of non-fatal disease burden worldwide is skin disease.
most prevalent diseases worldwide are skin conditions.
Sources: World Health Organization (2023) · IHME, The Global Burden of Skin Disease
A dermatology assistant at the front line of care.
DermaScan AI turns a skin image into a ranked diagnosis and a clear next step, so the professional in front of the patient can act with confidence.
For clinics
A second opinion for general practitioners and nurses during every skin consultation.
- Early flagging of suspected melanoma and carcinoma
- Fewer unnecessary specialist referrals for benign lesions
- Faster, more consistent first assessments
For pharmacies
Help pharmacists know when a skin complaint is safe to manage at the counter — and when it needs a doctor.
- Clear "see a doctor" alerts for suspicious lesions
- Safer advice for common, benign conditions
- A trusted, differentiated health service for customers
Built on Skin-DeepNet: a four-stage deep learning pipeline.
DermaScan AI is based on Skin-DeepNet, a framework for automated early diagnosis and classification of skin cancer lesions in dermoscopy images, published in Scientific Reports (Nature Portfolio) in 2025.
-
1
Enhance & clean
Image contrast is boosted adaptively, and hair is detected and digitally removed so it doesn't hide the lesion.
AGCWDBlackhatInpainting -
2
Segment the lesion
A Mask R-CNN network finds the lesion, and GrabCut refines its exact border, separating it from healthy skin.
Mask R-CNNResNet50 + FPNGrabCut -
3
Extract features
A high-resolution network with an attention block learns fine visual patterns, analysed by two parallel deep classifiers.
HRNet-W32AttentionDRBMDBN -
4
Fuse the decision
The classifiers' outputs are combined by an ensemble model to produce one final, more reliable diagnosis.
XGBoostStacking
Near-perfect accuracy on two benchmark datasets.
Evaluated on the public ISIC 2019 and HAM10000 dermoscopy datasets, using a 70% / 15% / 15% split for training, validation and testing.
| Dataset | Accuracy | Precision | Recall | F1-score | AUC |
|---|---|---|---|---|---|
| ISIC 2019 · 8 classes | 99.65% | 99.51% | 99.56% | 99.54% | 99.94% |
| HAM10000 · 7 classes | 100% | 99.92% | 100% | 99.96% | 99.97% |
Backbone comparison (ISIC 2019, before fusion)
| Feature extractor | Accuracy |
|---|---|
| ResNet101 | 95.94% |
| EfficientNetB0 | 96.74% |
| HRNet (selected) | 97.54% |
Compared with published methods (ISIC 2019)
| Method | Accuracy | F1 |
|---|---|---|
| Singh et al. | 98.04% | 96.24% |
| Radhika & Chandana | 98.77% | 98.76% |
| Skin-DeepNet | 99.65% | 99.54% |
All figures from Al-Waisy et al., Scientific Reports 15, 31234 (2025). doi:10.1038/s41598-025-15655-9
Eight types of skin lesion, malignant and benign.
The ISIC 2019 classes and the number of dermoscopy images of each used to develop the model.
Total: 25,331 images. Rare classes were balanced during training with shifts, flips and rotations.
Honest about what the model can and can't do.
DermaScan AI supports clinical judgement — it never replaces it. These are the known limits we design around.
Rare classes are harder
On ISIC 2019, recall was 79.2% for squamous cell carcinoma and 81.5% for vascular lesions, the least represented classes — so suspicious results always trigger referral.
Benchmark-tested, clinic-next
Results come from public benchmark datasets. Validation on real-world patient images from clinics and pharmacies is the next step.
The professional decides
Every result is decision support. Diagnosis and treatment always remain with a qualified healthcare professional.
Where computer science meets bedside care.
DermaScan AI was founded in 2025 by a research team combining computer science and nursing — machine-learning expertise and first-hand clinical experience in one product.
Computer Science
Faculty of Science. Builds the deep learning model, the image-processing pipeline and the diagnosis application.
Nursing
Brings clinical knowledge of patient care, triage and referral, so results fit real consultations in clinics and pharmacies.
Frequently asked questions
Does DermaScan AI replace a dermatologist?
No. It gives clinicians and pharmacists a fast first assessment and flags suspicious lesions that need a specialist. The final diagnosis always belongs to a healthcare professional.
What kind of images does it analyse?
The model was developed on dermoscopy images — magnified, well-lit images of skin lesions — from the ISIC 2019 and HAM10000 datasets.
Which skin conditions can it identify?
Eight lesion types: melanoma, basal cell carcinoma, squamous cell carcinoma, actinic keratosis, melanocytic nevus, benign keratosis, vascular lesions and dermatofibroma.
How accurate is it?
The underlying Skin-DeepNet framework reached 99.65% accuracy and 99.94% AUC on ISIC 2019, and 100% accuracy on HAM10000, as published in Scientific Reports (2025).
How does it work?
Four stages: the image is enhanced and cleaned of hair, the lesion is segmented, deep features are extracted with an HRNet network and attention block, and an ensemble model fuses the results into one diagnosis.
Peer-reviewed science behind every scan.
Read the full open-access study behind DermaScan AI's diagnosis engine.
Al-Waisy, A. S., Al-Fahdawi, S., Khalaf, M. I., Mohammed, M. A., Al-Attar, B., et al. (2025). A deep learning framework for automated early diagnosis and classification of skin cancer lesions in dermoscopy images. Scientific Reports, 15, 31234.
https://doi.org/10.1038/s41598-025-15655-9