zikabyte/skin-problem-detection-demo

Skin Problem Detection (YOLO26)

Object-detection model that localizes common facial-skin concerns. Trained as a technical feasibility test, not a production/medical product.

• Base model: yolo26n.pt
• Framework: Ultralytics YOLO26
• Image size: 640
• Epochs completed: 60
• Dataset: skin-problem-4f8bc (CC BY 4.0), 9182 images

Classes (4)

• 0: Mole
• 1: acne
• 2: rosacea
• 3: wrinkle

Metrics (validation split, best checkpoint)

Precision	Recall	mAP@50	mAP@50-95
0.543	0.5102	0.4955	0.273

Held-out test split

Overall: precision 0.5523, recall 0.5308, mAP@50 0.5211, mAP@50-95 0.2897

Class	mAP@50-95
Mole	0.6638
acne	0.0989
rosacea	0.2436
wrinkle	0.1525

Per-class performance is uneven: well-defined lesions score far higher than small, clustered, or diffuse ones. See metrics_test.json.

Usage

from ultralytics import YOLO

model = YOLO("best.pt")            # or "model.onnx"
results = model.predict("face.jpg", conf=0.25)
results[0].show()

Limitations

Not a diagnostic tool. Trained on a single-source dataset of ~700px images; expect domain shift on different lighting, skin tones, and camera setups.