Currently, I'm trying to optimize a training process of a neural net to improve final results. The problem I'm dealing with is multiclass segmentation on microscopic data.
The paradox is that the best (and not sufficient) result is giving the simplest U-Net architecture on the original dataset. If I try a deeper or more complex model (e.g. r2unet), the final segmentation is significantly worse. If I try on the fly augmentation - worse as well. Changing a complex model into a more shallow one didn't help either (just tried out the other way than making it more complex).
Now, I'm trying to make a custom loss function work to improve the segmentation.
Any ideas what might be a root cause? Or any other ideas that could improve the result?
To get more specific, here's an example of the data. The initial 4000x4000 images are cut to 512x512, which results in a little over 3700 images. Most of them don't include the classes and is just background, that's why I'm trying to make another loss functions work, as well as weighted classes.
So far, I'm using categorical cross-entropy as a loss function, however, dice, Jaccard, the focal losses seem could be more suitable and once I'll finish my computations I'll try to make these work again, so far my tries weren't really compatible with Keras, at least it seemed.
The size of U-Net
- depth = 5
- first conv layer has 64 filters, goes up to 1024.
R2U-Net
- depth 5
- first layer 64filter (tried also 32)
