I am doing a brain MRI segmentation task using Unet model. My problem is it does no matter what transformation i use during training my val dice score after few epochs stucks around 77% score.
I already have a successfuly trained model with 93% dice score on my test set, but i wanted to improve it by adding some augmentation to my training. I used transformations from the albumentation library. I only experimented with transformations that are reveleant to my train case. (Horizontal, VerticalFlip, etc ..) Things that i already tried:
- waiting for more epochs
- using different transformations, using only one
- using different encoders to make the model more deep, (resnet34, resnet152 etc)
