ConvexAdam and FireANTs

Overview

Two optimization-based registration methods are wrapped behind the MIR interface for baselines and for hybrid pipelines:

• ConvexAdam (Siebert et al.): discrete convex optimization on hand-crafted MIND features followed by Adam-based instance optimization. MIR provides the MIND variant, SVF and spatially-varying-regularization variants, and versions that optimize on learned features (nnU-Net segmentation features, VFA features).

• FireANTs (Jena et al.): GPU implementations of greedy and SyN diffeomorphic registration, with SPR variants, under MIR.models.fireants.

Usage

from MIR.models import convex_adam_MIND, convex_adam_MIND_SPR, get_ConvexAdam_MIND_brain_default_config

config = get_ConvexAdam_MIND_brain_default_config()
flow = convex_adam_MIND(mov, fix, config)                 # displacement in voxels

Other entry points: convex_adam_MIND_SVF, convex_adam_features / convex_adam_features_svf (precomputed features), convex_adam_seg_features, convex_adam_vfa. FireANTs: MIR.models.fireants.registration.greedy / syn and their _SPR counterparts.

Key modules

• MIR.models.convexAdam (convex_adam_MIND, configs_ConvexAdam_MIND, convex_adam_VFA, …)

• MIR.models.fireants

Related tutorials

• LUMIR25 ConvexAdam

• LUMIR24 ConvexAdam

• autoPET ConvexAdam-SPR

• autoPET FireANTs-SPR

• Brain registration (ConvexAdam-MIND)

References

• Siebert, Grossbroehmer, Hansen, Heinrich. ConvexAdam: self-configuring dual-optimization-based 3D multitask medical image registration. IEEE TMI 2024. Paper | Repo

• Jena et al. FireANTs: adaptive Riemannian optimization for multi-scale diffeomorphic registration. Repo