How should I reproduce the training process of Swin UNETR on BTCV so that it can achieve the mean dice score recorded in the paper?

[JeMing-creater]

Describe the bug
Using multi-GPU training from scratch, the BTCV segmentation score of 91.8% the average dice score cannot be achieved.

To Reproduce
Steps to reproduce the behavior:

1. Parallel training is performed on 4 T100s.
2. The batch size is set to 1, and max_epochs is set to 5000. Other settings are selected as default.

My confusion is:
I cannot reproduce the average dice score of 91.8%, and the highest is only 83%. What training method should be used to achieve the scores recorded in the paper?

[ASCI02]

Hi, have you solved this problem

[litleWY]

Have you reproduce the results now?

[ASCI02]

No

…

---Original--- From: "wendi ***@***.***> Date: Thu, May 8, 2025 13:38 PM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [Project-MONAI/research-contributions] How should I reproduce thetraining process of Swin UNETR on BTCV so that it can achieve the mean dicescore recorded in the paper? (Issue #302) litleWY left a comment (Project-MONAI/research-contributions#302) Have you reproduce the results now? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[litleWY]

I think the reason our reproduced Dice score is 0.83 instead of 0.918 is that we didn’t use five-fold cross-validation, which the authors mentioned in the paper. But I haven’t tried this method yet. You can try it.