- Project
- AI/ML
- Research
- Solo
- 2026
Skin lesion classification (HAM10000)
Three modern vision backbones compared on 7 imbalanced classes, judged by macro-F1.
How it works
- 01HAM10000: 10,015 dermoscopy images, 7 classes
- 02ConvNeXt V2-Tiny, DeiT III-Small, MaxViT-Tiny
- 03Macro-F1 as the primary metric
ConvNeXt V2 macro-F1Up from 0.635; AUROC 0.957 on 7 imbalanced classes.
ConvNeXt V2 macro-F1
Up from 0.635; AUROC 0.957 on 7 imbalanced classes.
The problem
In skin lesion data most images are benign, so accuracy hides how badly rare classes are handled.
My part
Ran the comparison and the improvements on ConvNeXt V2.
What did not work
- The first ConvNeXt V2 run reached 0.635 macro-F1.
Limitations
- A course project on a public dataset; not validated clinically.