TODO

publish brain model

high resolution brain model

eval lncc brain model

dice for brain model

dice for ants for brain model

replicate synthmorph results on brain by downloading code

trim gradicon knee model

add reloading script

merge doc pull request

update google doc merge fix pull request

Evaluation catastrophe

ICON results at half resolution:

Old result (full resolution segmentation, merged segmentation): 68.3

Full resolution segmentation, seperate cartilages: 66.5

Half resolution segmentation, seperate cartilages

GradICON results at half resolution:

Old result: (full resolution segmentation, merged segmentation): 71.3

Full resolution segmentation, seperate cartilages: 70.05

Half resolution segmentation, seperate cartilages: 68.4

GradICON results at full resolution:

Old result (merged cartilages): 74.0 (73.3 reported in paper)

Full resolution segmentation, seperate cartilages: 72.8

Note:

The GradICON half resolution knee model is undertrained: the ICON result has two steps at quarter resolution, two steps at half resolution. The GradICON full resolution is two steps at quarter, one step at half,k one step at full (full resolution replaced one of the half resolution steps) and the GradICON half resolution model is just a training artifact from midway through training the full resolution model. As a result, if 68.3 isn't impressive enough, I can almost certainly boost it by adding a step to match the ICON architecture.

Brain model

I have two brain models available now: half resolution LNCC GradICON, and half resolution MSE GradICON.