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
155 stars 22 forks source link

Dice Score is low #20

Open zzh1230 opened 1 year ago

zzh1230 commented 1 year ago

I trained on the LiTS data set, and the Dice Score on the training set can reach 0.9035, but only 0.6898 on the test set. When I visualized the results, I found an interesting thing. The predicted shapes are very similar, but there seems to be some Offset, where might the problem be? image

920232796 commented 1 year ago

I do not encounter your problem before, it is very interesting. You can check if there are problems about resample or crop non-zero operation in the data loading processing.

OCEANOUXIN commented 1 year ago

Oh, I met the same problem with you! T^T