Overlapping segmentation using png input and png label

[NaJiwoong]

Hi,
I read some documents and issues in your repository that explains the region based segmentation for segmentation of overlapping label.
However, I think somewhere I've done it wrong, so I want to ask you about the json file and png.
For example, current my state is like below.

So I want to make it possible to get segmentation results of "A" as ellipse region, "B" as the triangle regions, "C" as the rectangular region.
So if I understood correctly, I think the json file should be as :
...
"labels": {
"background": 0,
"A": [1, 2, 3, 4],
"B": [2, 4],
"D": [3, 4],
"E": [4] (or 4)
},
"regions_class_order": [1, 2, 3, 4],
...
Is it right?

Also, how should I make png file?
Currently, I've tried two cases:

1. Just colorize the regions 1, 2, 3, 4 as arbitrary different colors.
2. Make the RGB value of the region as (1, 1, 1), (2, 2, 2), (3, 3, 3), (4, 4, 4) in png file.

I thought if I make a lable as 1, then I think the framework cannot tell which region corresponds to which label.
So I also made a label as 2, even though I cannot distinguish the region in my eye (in 0~255 scale), but the values can be distinguished by the nnUnet code.

Also, I set followed the instruction "Basically you can define each of your labels as a separate binary classification with a sigmoid function + Dice&BCE loss (instead of softmax + Dice&CE loss)"

However, I think the training process goes wrong. (the value 4 covers almost predicted mask)
Are there something I've missed?