Marinheiro2004/UNet-CAA

UNet Kvasir-SEG

Architecture

• Base: U-Net
• Encoder: ResNet34 (pretrained)
• Input: 3 channels, 512×512
• Output: 1 channel

Loss function

• Dice Loss

Dataset

• Kvasir-SEG
• Train/Val/Test: dataset default split

Preprocessing

• Resize 512×512
• Data augmentation: HorizontalFlip, VerticalFlip, RandomRotate90, RandomBrightnessContrast

Notes

• Model trained with 100 epochs
• Best checkpoint based on Dice Loss Validation

Config

{
  "in_channels": 3,
  "num_classes": 1,
  "encoder": "resnet34",
  "img_size": 512
}

Performance

• Best validation Dice Loss: 0.122
• Corresponding Dice Score ≈ 0.878