Medical Image Segmentation with U-Net: Reaching Dice 0.7964

Architecture: U-Net with EfficientNet Encoder

import segmentation_models_pytorch as smp

model = smp.Unet(
    encoder_name='efficientnet-b4',
    encoder_weights='imagenet',
    in_channels=3,
    classes=1,
)

Loss Function: Combo Loss

Dice + BCE combination outperformed either alone:

loss = 0.5 * bce_loss + 0.5 * (1 - dice_score)

Augmentation Stack (Albumentations)

• RandomResizedCrop, HorizontalFlip, VerticalFlip
• ElasticTransform, GridDistortion
• CLAHE, RandomBrightness
• CoarseDropout (Cutout)

Post-processing

Test-time augmentation (TTA) with 8 flips/rotations added +0.015 Dice.