guo-stone / mambamorph Goto Github PK

感谢作者的工作贡献，我遇到了一个问题，我在训练完成VoxelMorph、TransMorph、mm-feat后，进行test，3个模型的DICE结果均为62.42±3.29，这让我很困惑，也就是说形变场没有形变能力，我打印了测试中的扭曲后的标签(warped_label)和原标签(src_label)进行比较，结果显示扭曲后的标签(warped_label)相较于原标签(src_label)并没有改变。随后我在训练阶段的每一个step中都打印这个扭曲后的标签和原标签进行比较，结果都是不变的。这让我很惊讶，今天，我把学习率调整到了运来的40倍(原来1e-4，现在4e-3)，结果才会变。我不知道我遇到了什么问题，所以向作者询问，谢谢您！

'/root/MambaMorph-main/voxelmorph/init.py'代码里面缺少voxelmorph文件夹里面的模块

Traceback (most recent call last):
File "/home/lx/dl/reg/mamba/./scripts/torch/train_cross.py", line 64, in
import generators as src_generators
ModuleNotFoundError: No module named 'generators'

需要在Linux系统里下载好mamba才能把代码跑起来吗

Thank you for your contribution to this work. I have a question: by incorporating the DICE score between labels as part of the loss function, does this imply that the network is a semi-supervised registration network rather than a traditional unsupervised registration network? I would be grateful for your response.

dear authors
It's really a meaningful job. I want to as where can I get the CT-MR registration SR-Reg dataset, the link in your paper seems not accessible.

作者您好，非常感谢您的工作，我在复现您的工作的时候出现维度无法拼接的问题如上图所示，可以问一下您是如何解决的吗？