output window size of 3D U-net

[niftk]

output window size of 3D U-net

[wyli]

Hi @nguyenhg,

Why image_size should be divisible by 8 and >89?

so that we could handle multiple layers of "2×2×2 downsampling/upsampling with strides of two" conveniently.

The parameter spatial_window_size of [label] should be same with the one of [images] or not?

Not for the 3D-Unet, please check out the original release for more info: https://lmb.informatik.uni-freiburg.de/resources/opensource/unet.en.html
and the script for image/feature dims: https://lmb.informatik.uni-freiburg.de/resources/opensource/findsize.m

[wyli]

Hi @nguyenhg thanks for the feedback. This is using the 3D U-net (https://arxiv.org/abs/1606.06650) and by the original design, it takes 132×132×116-voxel inputs and predicts 44×44×28-voxel outputs. Both of your Q1 and Q2 are due to this "cropping" effect of the network (also please note there's another spatial_window_size parameter in the [inference] section of the config).

You could also remove the cropping layer by deleting this line https://github.com/NifTK/NiftyNet/blob/dev/niftynet/network/unet.py#L122
the result is a network taking NxNxN inputs and predicting NxNxN outputs. Hope this helps!

[enzo-cin]

Hi @wyli, thank you for your comment. I did the interpolation for my data to original design: 132×132×116 for both images and labels, voxel size is 1.76×1.76×2.04 . And re-run the training step, but the dimension is not compatible.
Could you explain me:
Why image_size should be divisible by 8 and >89?
The parameter spatial_window_size of [label] should be same with the one of [images] or not?

[YilinLiu97]

Hi@wyli, with the cropped version, the border has to be specified to 44, which makes the inference too long and results in bad performance. Do you have any suggestion on this? Thank you.

[NicoYuCN]

Does it mean the maximum size of input volumetric images is no larger than [132 132 116]? If so, it makes the NiftyNet is not so flexiable.