search

MIC-DKFZ / nnUNet

Apache License 2.0
5.9k stars 1.76k forks source link

The mean validation dice is significantly lower than the pseudo dice. #1995

Closed 111678900 closed 5 months ago

111678900 commented 8 months ago

The mean validation dice is significantly lower than the pseudo dice.Please help me

sun-wanqing commented 8 months ago

I encountered the same problem and validation dice was only half of pseudo dice.

111678900 commented 8 months ago

您好，我查出来是数据问题

---Original--- From: @.> Date: Wed, Mar 13, 2024 11:23 AM To: @.>; Cc: @.**@.>; Subject: Re: [MIC-DKFZ/nnUNet] The mean validation dice is significantly lowerthan the pseudo dice. (Issue #1995)

I encountered the same problem and validation dice was only half of pseudo dice.

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

111678900 commented 8 months ago

不知道您是否是相同问题

---Original--- From: @.> Date: Wed, Mar 13, 2024 11:23 AM To: @.>; Cc: @.**@.>; Subject: Re: [MIC-DKFZ/nnUNet] The mean validation dice is significantly lowerthan the pseudo dice. (Issue #1995)

I encountered the same problem and validation dice was only half of pseudo dice.

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

sun-wanqing commented 8 months ago

您好，我查出来是数据问题 … ---Original--- From: @.> Date: Wed, Mar 13, 2024 11:23 AM To: @.>; Cc: @.**@.>; Subject: Re: [MIC-DKFZ/nnUNet] The mean validation dice is significantly lowerthan the pseudo dice. (Issue #1995) I encountered the same problem and validation dice was only half of pseudo dice. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

我应该不是欸，请问您现在这两个dice差别还大吗？但两个计算方式感觉是不一样的，一个是batch算的，一个是单个样本算好求平均的。

111678900 commented 8 months ago

我的是数据集有几张不太好，大部分效果很好

sten2lu commented 6 months ago

nnU-Net is trained with patches of the images which are samples from the original images. These patches are also used to calculate the pseudo-dice metric during training.

After training, nnU-Net performs sliding window inference on the entire image. The patches obtained from this sliding window sampling may differ from the sampling strategy used during training, potentially decreasing the validation-Dice score.

You could check for the background class having a low DICE score, as nnU-Net uses an oversampling strategy for foreground classes.