Why do u use binary_dilation for boundary 1-8?

[Bonsen]

I found the pixel size of contour is 4 when disk equals to 2, and it will cover pixels (==2) inside nucleis and outside nucleis (==2). And if contours are thick，the prediction covers more pixels inside the nucleis. What's more, if they are overlapped, the pixels outside this nucleis may cover some pixels inside other nucleis. It's hard to discriminate them.
Or u just want to generate boundary for test?
Do u use 2pixels boundary to train finally?