Increase dice score

vision
1

Hi all, please be patient to me since I’m new to semantic segmentation.

I’m implementing U-Net based on its original paper. I’m using DiceCELoss from MONAI. After 70 epochs with batch_size=30 I got this result:

Train Loss
Train Loss782×674 38 KB

Val Dice
Val Dice746×682 26.7 KB

Val Loss
Val Loss770×678 26.6 KB

From the graphs I can tell that, the train loss is kind of fluctuated around, however, the validation loss and dice score seem to be stable, and not gonna be improved soon.

Is there a way to solve this problem? Or is there any way that I can improve my dice score to around above 90%?

Thank you!

2

Typical things to try:

  • Change model arch (width / depth), why does your model not overfit the data?
  • Look at your results, find out something about the failure and see if there are systematic things your model gets wrong.
  • One black box thing that might help is learning rate schedules (eg decrease lr by multiplying with 1/3 after, say, 40 epochs).

Best regards

Thomas

1 Like