AlexNet and CIFAR analogs in MRI data

, ,
1

hi guys,

  1. the question is: is there any canonical dataset for pathology classification on structural mri and functional mri with 200> subjects and unprocessed data?

  2. for that dataset if there is any canonical 2D (or most probably 3D) CNN with optimized architecture? to have any stated result to compare with?

  • ABIDE and OASIS seems to be ancient and ABIDE with soa = 0.68 something.
  • HCP has no pathology to classify
  • SchizConnect data from different sources, so problems with domain adaptation
  • openneuro small datasets

and with the model architecture for 3D CNN, we experimenting more than a year with inhouse and openneuro data, but there is no architecture, except of canonical VoxResNet of VoxCNN to reference

our solution so far is:
https://ieeexplore.ieee.org/document/8637478

do you have any thoughts in mind about that?

1 Like
2

We have been developing the nobrainer package to provide baselines networks that people can reuse.

from a model perspective, only one network is available, but hope to see that change over the next few weeks.

once we have these stable anatomical networks, they can be used for transfer learning to smaller datasets.

1 Like
3

thank you! I found only segmentation models (unit) in the repository, have you tried any classification so far?

by the way, we were using neurodocker to install ANTS, and thanks for that :slight_smile: my appreciation for the developers

2 Likes