I'm trying to build a model to segment brain tumors.
I trained a model, and the validation dice coefficient is disappointing(0.6). When i saw the predicted images with the ground truths, it seems like the model can't segment tumors well in unusual images like the right of the below. The left image is an usual brain tumor image. (please note the color of the tumor)
And i'm not sure, but the dice score seems to be related with the input range of tumor region.
low dices - 0 centered input
high dices - about 3 centered input.
The input were standardized.
1. How does i make the model think of 0 centered input as useful information?
2. Does it worth to try other normalization methods(e.g. min-max scale) in this case?
3. In this case, what methods are helpful to improve the performance?