Problems about the landmark distance in "Recursive Cascaded Networks for Unsupervised Medical Image Registration

[lajipeng]

Dear Zhao:
Recently, I was using the dataset Sliver provided by you for learning-based image registration. Each scan in Sliver dataset contains six anatomical landmark and I have seen the landmard score is about 10 mm in your paper "Recursive Cascaded Networks for Unsupervised Medical Image Registration" . However, I have a few questions.
1、 What do these landmarks mark? Is there a specific standard？
2、It seems a landmark error >12 mm shows that the alignment inside of liver is not perfect yet because the Lm dist < 2 mm in lung CT registration (e.g DIR lab). Since your model works so well, I don't think that's the reason for your approach, but I'm curious about the reasons for this bias.
I didn't find any discussion of this in your paper, but using your data set made me have to look at these issues, and could you help me to answers these questions ? I would appreciate your answers very much. Meanwhile, thank you very much for your excellent work and open source dataset since I have learned a lot from your research.

[zsyzzsoft]

Hi, 1. You may refer to the VTN paper, "Unsupervised 3D End-to-End Medical Image Registration with Volume Tweening Network", which includes the standard for landmark annotation. 2. Our landmark distance is measured in pixels rather than mm so they are not directly comparable. Best, Shengyu Alan ***@***.***> 于 2021年7月30日周五 15:17写道：

…

Dear Zhao: Recently, I was using the dataset Sliver provided by you for learning-based image registration. Each scan in Sliver dataset contains six anatomical landmark and I have seen the landmard score is about 10 mm in your paper "Recursive Cascaded Networks for Unsupervised Medical Image Registration" . However, I have a few questions. 1、 What do these landmarks mark? Is there a specific standard？ 2、It seems a landmark error >12 mm shows that the alignment inside of liver is not perfect yet because the Lm dist < 2 mm in lung CT registration (e.g DIR lab). Since your model works so well, I don't think that's the reason for your approach, but I'm curious about the reasons for this bias. I didn't find any discussion of this in your paper, but using your data set made me have to look at these issues, and could you help me to answers these questions ? I would appreciate your answers very much. Meanwhile, thank you very much for your excellent work and open source dataset since I have learned a lot from your research. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#59>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFMGUM5Z6MK22SFGMIWKWF3T2JGYRANCNFSM5BH6M2WA> .

[lajipeng]

Thank you very much. But the VTN paper only contains 4 points. In the Sliver dataset, there are six points.
Moreover, I feel it's difficult to lower the Lm dist. is it because the two volumes are different a lot. How can this be improved?

[zsyzzsoft]

The two extra points are labeled after VTN published, and I cannot find their description for now. This metric is indeed not an easy metric as every scan comes from a different person, and the model needs to understand the key points beyond similarity to improve this metric.