GradICON medium resolution model released: pull request

Update iconregistration pip install iconregistration==0.3.3 and switch to the new model #model = pretrainedmodels.OAIkneesregistrationmodel() model = pretrainedmodels.OAIkneesgradICONmodel() This release runs at half resolution just like the old model, with much fewer folds, better dice, and faster runtime: action with results

Let me know if you also want the full resolution model: it's not quite ready for release right now because it only runs on gpus with at least 20 GB of vram


Ideas attempted:

Focus registration on lung instead of ribs

  • only evaluate loss in lung region of moving image

  • only evaluate loss in lung region of fixed image

  • only evaluate loss in intersection of lung segmentations

  • Zero image intensity outside of lung region

  • Clip image intensity to -1000, 0 before training ✅

Combat overfitting

  • Warp both fixed and moving image with same linear transform

  • Train on longitudinal as well as cross-subject pairs

  • Warp both fixed and moving image with different linear transform ✅

Best (bad) result so far:

mTRE: 9.770269366339061, mTRE_X: 2.2851461818021503, mTRE_Y: 6.378642569335982, mTRE_Z: 5.522105794414662, DICE: 0.9608421929940194
mTRE: 7.972476933755549, mTRE_X: 2.5015854799159594, mTRE_Y: 4.762276590067334, mTRE_Z: 4.4913444717735675, DICE: 0.9781910842486858

Ideas not tested yet:

  • LNCC

  • Bring Lin's architecture forward

  • Do highest resolution training

  • Do overfit



Task for marc: fix permissions on test set?

Training setup:

2 trainings currently in flight.

Both λ=.05\lambda = .05

Both skull stripped

one at half res, the other progressive starting with quarter res.


Only visual results so far- evaluate with me?


  • Evaluate on MindBoggle101

  • Evaluate on LPBA40

  • Just do the quicksilver evaluation

ICON library tutorial

Working on an ICON library tutorial here

Transformer compatible with GradICON loss?

Not as far as I can tell

Theory TODO

  • Hessian of loss

  • 1-D experiment

  • fewer or more samples

  • ICON without gradicon, but with loss barrier around J=0|J| = 0

  • optimization without network

Back to Reports