search

jadie1 / Point2SSM

Point2SSM: Learning Morphological Variations of Anatomies from Point Cloud
https://arxiv.org/abs/2305.14486
4 stars 0 forks source link

Multi anatomy training #2

Open liamburrows opened 3 weeks ago

liamburrows commented 3 weeks ago

Hi Great work! I have been running it on my own dataset and it works really well!

Just wondering about how to do the multi anatomy case? I note there's a small section in the appendix of the paper. Is it just a case of concatenating the input vectors to have (in the case of two anatomies) double the amount of input points, then having double the output points which you split the output vector into two so you have a vector for each anatomy?

Thanks!

jadie1 commented 2 weeks ago

Hi I am glad it is working well for you! For the naive way of multi-anatomy training we tried in the appendix, we simply created a new data folder with both anatomy types. So the training and validation folders just contain both types. The model does not have a way of distinguishing between the anatomy types but it effectively learns correspondence for both. So there is no concatenation, we just treat each shape as a separate instance disregarding the type. Let me know if that makes sense. You can also check out our most recent work Point2SSM++ where we proposed adding an auxiliary anatomy classifier to the network. This additional training signal helps the model capture anatomy specific features.