search

ge-xing / Diff-UNet

Diff-UNet: A Diffusion Embedded Network for Volumetric Segmentation. (using diffusion for 3D medical image segmentation)
Apache License 2.0
149 stars 21 forks source link

unable to reproduce the results of BTCV, only a mean dice score of 0.2. #39

Open XinYue-Yu1031 opened 2 months ago

XinYue-Yu1031 commented 2 months ago

Hello,

I trained BTCV on a single V100 GPU with a random data split ratio of 4:1, using a batch size of 1 and training for 3000 epochs. However, my mean dice score is only 0.2. Do you have any good training suggestions, or could you provide your BTCV training model? I would greatly appreciate it. 1 2 3

ge-xing commented 2 months ago

The training loss is very strange. I think you should try the BraTS dataset and see if there are still any problems.