Friday status

Monai work

It's almost completely done. There is a small discrepancy between the original and MONAI verions of the library when loading models trained with the old version of the library.

Here is a difference image between a warped brain registered with the original library, and a warped brain registered with the monai version:


The images range from values 0-1, and the errors are in the range 0.03. The change to the deformation field is subpixel but it should be zero.

In addition, there is also a small discrepancy in lung registration.

Compared to the original library, the test pair has errors that are larger by approximately a .01 mm.

TODO the test case for TwoStepConsistent is missing. Also, the ConstrICON models were never formally published and need to be- currently just an anonymous google drive link referenced in a colab notebook I think.

Equivariance Experiments:

We aim to try:

  • Weaker 2d feature extraction

  • Try retina using 3d code

  • 3D circles and squares

  • White noise to 2d case

We want to reproduce the effect where the equivariant net only captures the translation, and all deformation is left to the subsequent net.

Weaker 2d feature extraction

  • In 2d, we drop the power of the equivariant network from 128 channels to 12 channels, both in the encode and in the transformer. This harms performance, but not in the form of a reversion to translation


3D circles and Triangles



Back to Reports