I am a graduate student working on an conditional multi-organ segmentation model!
I have finished writing my code to train and validate segmentation architectures (like U-Net, PIPO-Net, DFCN). Currently I am trying to improve my multi-segmentation net with a Body Part Regressor (BPR - much like this one https://arxiv.org/pdf/1707.03891.pdf) which maps a CT image-slice to its corresponding location in the body, i.e. outputs a score which represents the height in the body where the input slice is positioned.
So my BPR can tell me what body part/area my volume (or sub-volume) shows. This means I could use partial evidence for my segmentation net, much like telling the net: “don´t look for liver or kidney while segmentating, since the BPR says the volume is only in the upper lung area”. But I do not want to set those partial labels manually.
Likewise, I could maybe combine the BPR and the segmentation net while training, making this partial evidence a learnable parameter.
Does somebody have an idea on how to combine the segmentation masks and the BPR scores or knows how to make a single architecture of those two?
Thanks for the ideas!