Loss Curves

Loss measures how well the network predicts masked pixels — lower is better.

The loss chart shows two curves

Training Loss (typically red/orange)

Error on the patches used for learning. Should decrease steadily as the model learns.

Validation Loss (typically blue/green)

Error on held-out patches not used for training. Shows how well the model generalizes.

What to look for

Healthy Training

Both curves decrease together, then flatten. A small gap between them is normal.

Overfitting

Training loss keeps decreasing but validation loss starts rising. The model is memorizing training data. Early stopping will typically catch this.

Underfitting

Both losses stay high and don't decrease much. Try longer training, larger model, or different learning rate.

Instability

Losses jump around erratically. Try reducing learning rate or increasing batch size.

Note: Unlike segmentation, there's no Dice score for denoising. Loss is the primary metric — visual inspection of results is important.