Kaiseem / DAR-UNet

[JBHI2022] A novel 3D unsupervised domain adaptation framework for cross-modality medical image segmentation
Apache License 2.0
40 stars 5 forks source link

关于最后测试结果 #20

Closed QuintinDong closed 6 months ago

QuintinDong commented 10 months ago

打扰作者,我按您的训练策略,由于显存原因,在分割网络dataloader中将图像和标签resize到128,并将batchsize改为1,其余没有任何改动,训练100eopch,然后最终的测试结果为Dice per class: 0.8173324 0.0039876397 0.24330574 0.55620456 Overall Dice: 0.4052075743675232 ASD per class: 2.367802277804098 26.726564945072344 7.202804602593032 4.151737284935796 Overall ASD: 10.112227277601319,可以看出第二类和第三类分割结果异常的差,请问您觉得这是什么导致的呢?

Kaiseem commented 10 months ago

如果我没理解错的话,你是把图像和标签缩小了4倍训练?你有在测试的时候进行同样的缩小吗?还有就是batch norm/group nom对batch size敏感,这个影响挺大