Loss IFA on BraTS?

[chjwhere]

Hi Boyun,

Noticing your code:

When handling BraTS, remove sc_loss.

    self.fe_loss = self.csco * cs_loss

and your paper:
It is crucial to emphasize that, due to the irregular size of tumors in the brain, our DDSP does not utilize L_IFA designed based on the similar relative volume size of anatomical structures in this task.

Loss IFA as a very important method in your work, was not applied in BraTS segmentation. Would it perform better or worse on BraTS if it is applied?
Thank you for your great work!

[chjwhere]

作者你好，你的DDSPSeg网络在心脏器官分割域迁移任务中通过IFA损失将源域与目标域特征进行对齐，从而达到从CT域至MR域的迁移。但是在脑肿瘤分割任务中，您将此IFA损失移除。我看代码和论文中，其余的损失并没有将T2域和T1域(e.g.)的相关信息进行对齐，只有将latent_a和latent_a_r进行对齐；将pred_mask_a、pred_mask_a_r、label进行对齐；以及将pred_mask_b和pred_mask_b_r对齐。所以我不太明白，在脑肿瘤分割任务中，您是怎么将不同模态的两域对齐实现域迁移的目的。
感谢你的工作，非常期待你的回复！

[BoyunZheng31]

Thank you for your thoughtful question. In our work, the design of Loss IFA (likely referring to a method tied to anatomical structure volume consistency) was intentionally omitted due to the irregular and heterogeneous size of brain tumors, which fundamentally deviates from the assumptions required for such a method. While we did experiment with integrating IFA-like components during early stages, the results showed no significant improvement in performance on BraTS segmentation tasks. This aligns with the inherent challenge of brain tumors: their unpredictable morphology disrupts the spatial regularity that IFA typically relies on.

Our approach instead focuses on dual-domain generalization and semantic constraints, enabling the model to adapt to diverse distributions. This design choice is validated through ablation studies in the paper, which demonstrate robust performance even in the absence of IFA.

Thank you again for your engagement with our work!

[okfine888]

您好，我最近也在研究这个代码。我目前在调试train.py文件，运行的时候，损失降低异常，可以跟您沟通交流一下吗？