the prediction mask of the 3D volume CT, each slice kept the same

[DISAPPEARED13]

Hi there. I use DeeplabV3+ for 3D volume CT, for segmentation. But in validation, some prediction mask volumes are strange, like each layer is the same as each other. How can I solve this problem? Thanks a lot!

[DISAPPEARED13]

I found that's because I use the resnet, which will reduce the resolution of the z-axis from 16(original) to 1. That's funny...

[DISAPPEARED13]

edit the resnet type from ***_os16 to ***_os8 can solve this problem, but let's see what's going on.

[DISAPPEARED13]

Hi there, I've met some new issue, that is, to reduce the times of down sampling in z-axis, I have to change some kernel size and stride and padding when defining the resnet. And then the z-axis didn't do any down sample so that I don't have to interpolate anything of the final output. But the evaluation of the sampled patches didn't improve well, just from 0.5 to 0.6.