Skip connections use concatenation (not addition), transferring fine-grained spatial features from each encoder level to the matching decoder level. This is the key difference from ResNet residual connections, which use addition for gradient flow within the same path.
Input images are resized to 256×256 RGB. Output is a (B, 1, 256, 256) logit map; apply sigmoid > 0.5 to get the binary mask.
Loss Function
Combined BCE + Dice loss:
L = BCE(logits, targets) + DiceLoss(logits, targets)
Why not standard cross-entropy alone?
Polyp segmentation has strong class imbalance — background pixels vastly outnumber polyp pixels. A model that predicts all background achieves high pixel accuracy but is clinically useless. BCE alone is dominated by the majority class.
Dice Loss directly optimizes the overlap between prediction and ground truth:
It is not affected by background dominance since it only measures the ratio of overlap to total predicted + true positives. Smooth term = 1e-6 for numerical stability.
BCE + Dice combines stable gradients (BCE) with overlap-aware optimization (Dice). This is the standard default for medical image segmentation.
The model performs well on medium-to-large polyps with clear boundaries. Small or subtle lesions are harder — expected behavior for a from-scratch U-Net without attention mechanisms.
References
Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI.
Jha, D., et al. (2020). Kvasir-SEG: A Segmented Polyp Dataset. MMM.
