Model Architecture

1. Segmentation – ResUNet++

ResUNet++ is a hybrid model combining:

• Residual connections (from ResNet) to improve gradient flow
• Attention gates to focus on lung regions
• Dense skip connections to preserve fine details
• Squeeze-and-Excitation blocks for adaptive feature recalibration

This makes it more effective for lung boundary segmentation compared to traditional U-Net.

---

2. Classification – EfficientNet-B0

• Lightweight CNN architecture optimized for efficiency and accuracy.
• Pretrained on ImageNet and fine-tuned for TB classification.
• Final output layer: sigmoid activation → probability of TB-positive case.

---

3. Loss Functions

Task	Loss Function
Segmentation	Dice Loss
Classification	Binary Cross Entropy