1. U-Net
Motivation
- Small amount of training data in biomedical AI due to private information problems and expert-needed labeling
- The need for very precise segmentation
Architecture
- Contracting Path
- 3x3 conv + BN + ReLU
- 2x2 Maxpooling (stride = 2)
- Expanding Path
- 2x2 transposed conv
- concatenate with feature map from contracting path
- 3x3 conv + BN + ReLU

Techniques
- Data Augmentation
- Random Elastic deformation → increases model invariance and robustness
- Pixel-wise loss weight
- Weight map with more weights on the edge pixels

Shortcomings
- Fixed depth(4)
- Best performance is not guaranteed for every dataset
- Very computationally inefficient to find best depth
- Simple skipped connection
- Connects Encoder and Decoder that has same depth
2. U-Net++
Motivation
Made to overcome U-Net shortcomings
- Variety of U-Nets with different depths that share the Encoder

