kaczmarj / ProjectManagement

Project management

Creative Commons Attribution 4.0 International

0 stars 0 forks source link

[nobrainer] run quick evaluations on existing nets #2

Open kaczmarj opened 7 years ago

kaczmarj commented 7 years ago

Run the existing data through pre-trained 2D and 3D neural nets:

As part of this, download the niftis from brainbox to openmind. Save them to

/storage/gablab001/data/nobrainer/volumes/

kaczmarj commented 7 years ago

In terms of NiftyNet:

[ ] use all (many) of gablab's aparc+aseg files to train a niftynet model, and test on mindboggle 101 dataset, which contains manually labeled brains.
[ ] rewrite a niftynet model in pure keras/tensorflow.
[ ] test niftynet BRATS model on meningioma data.

kaczmarj commented 6 years ago

Update

Trained a nobrainer model (highres3dnet via niftynet) on openmind's T1+brainmask data (~4100 pairs of volumes). The T1s and brainmasks were freesurfer outputs. The T1s were likely interpolated to 1 mm iso. Tested the model on T1s from the unseen mindboggle-101 dataset. Dice coefficients 0.968 +/- 0.025 (mean +/- std dev). range is 0.749 to 0.985. median is 0.974.

See this video demonstrating the model's prediction on a T1 that it did not see during training.

kaczmarj commented 6 years ago

I am also tracking my work on this in a Google presentation.