Technical Summary
Multiclass Skin Lesion Classification
10,015 dermoscopic images across 7 lesion classes. Swin-T with ImageNet pretraining.
Approach
Swin Transformer for 7-class skin lesion classification. 88.7% accuracy with data augmentation.
Objective
Build Swin Transformer model for multi-class skin lesion classification.
Challenge
Class imbalance: rare lesions (DF, AKIEC) underperform. Visually similar categories.
Solution
Swin-T with multi-scale attention + data augmentation (rotation, flipping).
Dataset
HAM1k (10,015 images)
Classes
7 lesion types
Model
Swin Transformer Tiny
Functions
- Image preprocessing: 224×224 resize, normalization.
- Data augmentation: horizontal flip, rotation.
- Swin-T with ImageNet pretraining.
- Multi-class evaluation (P, R, F1).
Architecture
- Swin-T backbone for hierarchical features.
- Shifted window attention for efficiency.
- ImageNet transfer learning.
- Softmax for 7-class output.
- Cross-entropy loss.
Preprocessing
- RandomOverSampler for class balance.
- Resize to 224×224.
- Geometric augmentation.
- Photometric transforms.
- Pixel normalization.
Training
- Cross-entropy loss.
- Adam, LR=1e-4, batch=32.
- Validation monitoring.
- Accuracy, precision, recall metrics.
Class Performance
| Class | Precision | Recall | F1 |
|---|---|---|---|
| AKIEC | 0.71 | 0.65 | 0.68 |
| BCC | 0.82 | 0.84 | 0.83 |
| BKL | 0.79 | 0.76 | 0.77 |
| DF | 0.70 | 0.68 | 0.69 |
| MEL | 0.75 | 0.73 | 0.74 |
| NV | 0.94 | 0.96 | 0.95 |
| VASC | 0.88 | 0.85 | 0.86 |
88.7% with augmentation. Data enhancement improves rare class performance.