Anistropic vs. Re-sampled Isotropic

[rickymwalsh]

Hi Fabian,

Thanks a lot for this useful method and repo! I was curious about the choice of keeping the data anisotropic and using a combination of 2D/3D convolutions vs. re-sampling the data to isotropic resolution before training.

For my task of lesion segmentation in spinal cord MRI, where I'm currently using sagittal scans and nnU-Net method gives me 2.75mm x 0.58mm x 0.58mm spacing, I get better results by resampling all data to 0.5mm isotropic resolution and training a 3D U-Net. I believe this paper also found a similar pattern for different tasks (see Tables 3/4, comparing nnU-Net to "Linear", i.e. linear interpolation when re-sampling to isotropic): https://www.sciencedirect.com/science/article/pii/S0010482522004565

So I was wondering whether you observed some similar behaviour in your experiments? Did the factor of saving computational cost come into the choice of the anistropic 2D+3D setting or was it purely based on empirical results?

Thanks,
Ricky