RuntimeError: "length_2D" must be smaller or equal to image dimensions after resampling.

Hello,

So I get the following error when I am trying to process an image with version 4.0 of the code. This only occurs with the BF and SEM models.

File "/home/msavage/anaconda3/envs/axonDeepSegEnv1/lib/python3.8/site- packages/ivadomed/loader/mri2d_segmentation_dataset.py", line 152, in prepare_indices raise RuntimeError('"length_2D" must be smaller or equal to image dimensions after resampling.') RuntimeError: "length_2D" must be smaller or equal to image dimensions after resampling.

The image in question is 2160 x 1440 pixels at a resolution of 0.0064 micron per pixel.

The system running this package is Ubuntu 20.04

Any help would be greatly appreciated.

Best,

Michael

[GrimmSnark]

Thank you very much for all your advice, the labelling is working much better now!

[mathieuboudreau]

Thanks @GrimmSnark !

@mariehbourget and I played around a bit with some of the ADS settings on your dataset, and here are some reflections//results:

• Because of your datasets contrast, we saw better results using the TEM model.
• Using the TEM model instead of the SEM model changes our internal resolution during segmentation, so there may be less of a need to do padding to avoid the error you were getting with SEM.
• Using the TEM model with your true resolution of 0.0064 microns, we saw better results in segmenting large axons (at the cost of missing small axons):

• By using the TEM model but changing the input resolution in ADS to 0.0134 (which essentially, what it will be doing is “zooming” in the patches of the images during the segmentation to your image), we saw better results for small axons (at the cost of worse segmentations on your larger axons):

• Applying padding did not provide good results in our tests.

I think the consensus internally in our group is that if we were in your place, it would be better to use a resolution 0.0134 during the segmentation phase only in ADS (note that when creating morphology files, you should use your true resolution value to get accurate values), and then manually correct the segmentation of the large axons using the FSLeyes plugin we provide. Manually fixing/manually segmenting a few large axons is much less time consuming than fixing/manually segmenting many small axons, in our experience.

Hope some of this information is helpful!