Computer vision research
Skin images. A closer look.
Explore dermatoscopic images with machine learning.
1Add image2Run model3Explore scores
New analysis
Drop a dermatoscopic image here
Or choose a file from your device
JPG, PNG or WebP · up to 10 MB
Processed in memory. Never saved to an image history.
Preparing
Checking model availability…
Ready for a closer look
Add an image or try a dataset example below. Your class scores will appear here.
Highest-scoring class
Explore the reference collection
One example from each class. Labels come from the dataset.
Examples will appear once the dataset is ready.
A controlled experiment
Less fine-tuning. What changes?
Compare two MobileNetV2 configurations.
Baseline
4 feature blocks
Fine-tune the last four blocks and the classifier.
Our modification
2 feature blocks
Fine-tune the last two blocks and the classifier.
Evaluation protocol
Same data. Fair comparison.
Identical splits, settings and initial weights. Select the best epoch using validation macro-F1.
| Model | Accuracy | Macro-F1 | Best epoch | Training time |
|---|---|---|---|---|
| No results yet. Only measured values will be shown here. | ||||
Why go beyond accuracy?
HAM10000 is imbalanced. Macro-F1 gives each class equal weight.
A single run with a fixed seed is an educational experiment, not evidence of clinical effectiveness.
Computer vision research
Behind every score, there is an experiment.
DermaLab is an educational MVP for an ICT team project.