Deepreg: Demo + bug: unpaired brain MR pre-trained model performance poor

Created on 29 Aug 2020  路  5Comments  路  Source: DeepRegNet/DeepReg

Subject of the issue

The provided pre-trained network doesn't seem to warp the images fed in as provided in the demo.

If the documentation request is approved, would you be willing to submit a PR?

Yes / No

Your environment

  • DeepReg version (commit hash or tag) v. 0.1.6-alpha

  • OS (e.g. Ubuntu 20.04, MacOS 10.15, etc.) Mac High Sierra

  • Python Version (3.7, 3.8, etc.) 3.7

  • TensorFlow 2.3

Steps to reproduce

Run the demo as indicated for unpaired brain mri, and check the visualisation.

Expected behaviour

We would expect the pre-trained registration network to warp the moving image to resemble the fixed image.

Actual behaviour

As in Issue #347 the performance of the pre-trained model is strange - the pred-fixed-image (the warped moving image) and labels are not warped (appear the same), if we look at the visualisation.

聽聽Suggested fix

Train a model for longer...?

bug demo

Most helpful comment

I have now tried different hyperparameters and trained longer and the performance has improved. I am sharing a video example: https://liveuclac-my.sharepoint.com/鉁岋笍/g/personal/rmapcas_ucl_ac_uk/EaHqDbzAOAxKrPcsXAxnhkoBhuoGTs2NmYgAJ6K77XkLDQ?e=UgIuBc

How you made this video? I feel like we can add this feature;)

Honestly considering predict outputs all the images I don't see what's stopping us from writing a small script to make videos/gifs to visualise as a util for the package - simple file io...? I used some online gif generator but programatically easy?

Also @YipengHu I can send the video/gif for lung ct I made for the MEC and we can put it somewhere too?

All 5 comments

I have now tried different hyperparameters and trained longer and the performance has improved. I am sharing a video example: https://liveuclac-my.sharepoint.com/:v:/g/personal/rmapcas_ucl_ac_uk/EaHqDbzAOAxKrPcsXAxnhkoBhuoGTs2NmYgAJ6K77XkLDQ?e=UgIuBc

Nice - need a gallery for these

I have now tried different hyperparameters and trained longer and the performance has improved. I am sharing a video example: https://liveuclac-my.sharepoint.com/:v:/g/personal/rmapcas_ucl_ac_uk/EaHqDbzAOAxKrPcsXAxnhkoBhuoGTs2NmYgAJ6K77XkLDQ?e=UgIuBc

How you made this video? I feel like we can add this feature;)

I have now tried different hyperparameters and trained longer and the performance has improved. I am sharing a video example: https://liveuclac-my.sharepoint.com/鉁岋笍/g/personal/rmapcas_ucl_ac_uk/EaHqDbzAOAxKrPcsXAxnhkoBhuoGTs2NmYgAJ6K77XkLDQ?e=UgIuBc

How you made this video? I feel like we can add this feature;)

Honestly considering predict outputs all the images I don't see what's stopping us from writing a small script to make videos/gifs to visualise as a util for the package - simple file io...? I used some online gif generator but programatically easy?

Also @YipengHu I can send the video/gif for lung ct I made for the MEC and we can put it somewhere too?

I close this ticket since it's been improved by PR #422

Was this page helpful?
0 / 5 - 0 ratings

Related issues

mathpluscode picture mathpluscode  路  3Comments

mathpluscode picture mathpluscode  路  6Comments

s-sd picture s-sd  路  5Comments

mathpluscode picture mathpluscode  路  6Comments

mathpluscode picture mathpluscode  路  3Comments