Retraining accuracy not good

[AfrifaEben7]

Is there any way I can improve this accuracy?
I have a custom dataset (3D macrophages and cells) to segment. I adapted it to the BTCV dataset style and ran it as one would with the same approach. The best I am doing is a Dice score of 53%.

The metrics it starts with, that is what it ends with most of the time;