search

MIC-DKFZ / nnUNet

Apache License 2.0
5.79k stars 1.74k forks source link

Why is my dice coefficient constantly oscillating around 0 or 0.0x? #2200

Closed yibeizixia closed 4 months ago

yibeizixia commented 5 months ago

I tried running 25 lung ADC images (apparent diffusion coefficient maps), but the 3d_lowres failed, and the model's performance was poor. The dice coefficient always remains 0.0.

sten2lu commented 5 months ago

Hi @yibeizixia,

Could you send me the progress.png saved in the nnUNet_results for your trainings? Also, does the same issue appear for 3d_fullres and 2d?

So that I can gain a better overview of what might be going wrong.

Best regards,

Carsten

yibeizixia commented 5 months ago

@sten2lu Hello, thank you very much for your response. May I ask if I can send the results to your email after packaging them? Could you please provide your email address?

sten2lu commented 5 months ago

Hi, You can send the results per mail to carsten.lueth{at}dkfz.de.

yibeizixia commented 5 months ago

Hi,Carsten L, No problem at all, and thank you for reaching out. I'm a beginner in medical image segmentation myself.

This is my second time using the nnUNet model. The first time, I segmented 35 chest CT images and achieved excellent results. However, this time, while segmenting 25 CT images, I encountered two issues. Firstly, during the data preprocessing stage, I encountered the error message "3d_lowres not found in plans file nnUNetPlans.json of dataset Dataset029_Ours." Secondly, when using the 2D model, the dice coefficient consistently remained at 0.0 or slightly above, and the images kept oscillating. And,In my dataset.json, the "channel_names" field was written as "0": "CT". Previously, I had it as "0": "ADC". It seems like there's no difference between the two, as the dice score remains 0.0.

I'm not sure what the reasons could be, so I decided to reach out to you via email for assistance. Thank you for your help, and I wish you a pleasant day.

---- Replied Message ---- | From | Carsten @.> | | Date | 5/21/2024 15:38 | | To | @.> | | Cc | @.>, @.> | | Subject | Re: [MIC-DKFZ/nnUNet] Why is my dice coefficient constantly oscillating around 0 or 0.0x? (Issue #2200) |

Hi, You can send the results per mail to carsten.lueth{at}dkfz.de.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>

yibeizixia commented 5 months ago

@sten2lu Thank you. I've resolved the issue. It turned out that the insufficient data volume caused the phenomenon of getting 0.0-something pseudo Dice scores with only 25 ADC images. However, with over 50 images, the scores became normal.

sten2lu commented 4 months ago

This issue is already solved.