Blog post on ICON abstraction decisions in progress (Will include itkTransform ideas, will be presented to group on wednesday)


new training tends to overfit

Better train loss than ICON, worse test loss

evals to lowres DICS of [52]

Switch to all 22000 images

Improves test loss

evals to lowres DICE of 54.7

ICON achieved 57 at low resolution

however, alignment is visually better?

currently training high resolution second layer just to see if it works,

getting DICE 55

non-ATLAS ICON gets DICE 67 when used in atlas approach


New approach to patchwise registration


ToDO for next week

Train starting from ICON weights

augment data

Back to Reports
subdavis.com forrestli.com