Table 1 Model Parameters.
Task | Segmentation | Classification |
|---|---|---|
Model(s) | U-Net (VGG backbone) | ResNet-18, −34, −50 (ImageNet pretrained) |
Input size | 512 × 512 | 224 × 224 |
Batch size | 2 (frozen), 50 (unfrozen) | 32 |
Epochs | 100 | 100 |
Optimizer | Adam | Adam |
Learning rate | 1e-4 → 1e-6 | 0.0002 |
Scheduler | Cosine annealing | OneCycleLR |
Loss function(s) | Cross-entropy loss | Weighted BCE, Label smoothing (0.3) |
Regularization | / | Class weights for imbalance |
Data augmentation | Resize, random crop, horizontal flip, normalization | Random cropping, horizontal flip (p = 0.5), brightness/contrast ± 20%, normalization |
