search

MIC-DKFZ / anatomy_informed_DA

Apache License 2.0
18 stars 1 forks source link

Pretrained Model #1

Open farhancv09 opened 5 months ago

farhancv09 commented 5 months ago

Really interesting work. Can you provide the pretrained model for bladder and rectum segmentation?

Kobalt93 commented 5 months ago

Hi farhancv09, Thank you, I am happy that you find it interesting :) Unfortunately, our data protection does not allow us to make any "fingerprint" of the data including the trained models publicly available. Can I help you with anything else?

farhancv09 commented 5 months ago

Thank you for you prompt reply. And its really understandable what you said about data protection and fingerprinting. As are also dealing with similar challenges. I am thinking about in two ways this can be approached. 1) Maybe you can take a prediction on public dataset such as prostateX(https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=23691656) or PI-CAI (https://pi-cai.grand-challenge.org/) and provide the inference on some images. And we can use that inference as ground truth and train a model to get prediction on our in-house dataset (considering we don't need perfect organ segmentation). 2) If above is not feasible then, may you can provide more detail about manual segmentation process such as which tool, number of images etc. Please let me know whatever is feasible to you. One more question is, can't we use zones maps to for deformation and apply same transform to labels as well? or this will not produce plausible augmentations for prostate?

farhancv09 commented 3 months ago

Hello guys! seems like you guys are busy. If you got some time please respond to above question. Thank you.

Kobalt93 commented 2 months ago

Hi farhancv09, sorry for the late answer, the last period was busy and I was also out of the office for a longer time. In the meanwhile, TotalSegmentator for MRI is released where the supported classes set also includes the bladder and the entire colon: https://github.com/wasserth/TotalSegmentator I would start with that and check whether the predicted segmentations are feasible for your scans, they shouldn't be 100% accurate because of the calculation, but they should also have a good quality. If they wouldn't be good enough, you could manually correct them which is less work than annotating them from zero. For the annotation process, we use MITK: https://www.mitk.org/wiki/The_Medical_Imaging_Interaction_Toolkit_(MITK). What do you mean under zone maps? In the current setting, I also transform the labels, the ground truth is changing of course with the transformation.