No improvement pseudodice

[emlased]

Hello,

I am hoping to use nnUNet to segment coronary artery plaque. I have around 200 labeled studies with plaque segmented by anatomic region. I have tried each of the nnUNet trainers but I get output similar to below with no deviation from a pseudodice of 0.0 after > 50 epochs. I have also doubled checked the studies and ensured that they are labeled appropriately. Any help troubleshooting this would be greatly appreciated.

Here is my dataset.json file

{ "channel_names": {
"0": "CT" }, "labels": {
"background": 0, "Posterior descending": 1, "Right coronary": 2, "Left main coronary": 3, "Left anterior descending": 4, "Left circumflex": 5, "Mitral valve": 6, "Thoracic aorta": 7, "Aortic valve": 8 }, "numTraining": 201, "file_ending": ".nrrd" }

[mhxzkhl]

ME TOO In my train process the pseudo dice is irregular . I think this situation may be related to the patch size of the data

By the way, you should @ the person recommended by the author . Only then will there be a response

[emlased]

Thanks for the tip @mhxzkhl. I would really appreciate your input @saikat-roy

[saikat-roy]

Hi @emlased , have you found a solution to this issue?

[FWijanto]

Hi @emlased, @mhxzkhl , any chance you are trying to segment a really small structure? What's a typical ratio of volume to segment in the ground truth (in pixel) over total volume? Did you try the new residual encoder presets: https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/resenc_presets.md ?