DIRECT v2.2.0

DIRECT v2.2.0

Adaptive k-space sampling and optional motion registration, with the MIDL 2026 end-to-end ADS paper configs.

What's new

Adaptive dynamic sampling (ADS)
Learned 1D / 2D sampling policies (unified or frame-/phase-specific), wired through training and inference via additional_models.sampling_model.

Motion registration
Joint or disjoint recon↔registration (learned U-Net / VoxelMorph-style, or classical Demons / optical flow). Elastic deformations use an in-repo SciPy port — no elasticdeform dependency.

New models
MEDL and 3D VarNet paths used by the ADS setups, plus key-based losses and displacement-field logging for registration runs.

Paper projects

• projects/e2e_ads_recon — adaptive sampling + reconstruction (MIDL 2026)
• projects/e2e_ads_recon_reg — same pipeline with registration (arXiv:2411.18249)

Each config is a single YAML for both direct train and direct predict (inference block: 4× active, other rates commented).

Improvements

• inference.metrics: empty → reconstruct only; non-empty → score and write metrics_inference.json
• Safer config / checkpoint loading for paper and Hub-style YAMLs (unknown kwargs warned, optional training / validation)
• Calgary / CVPR configs cleaned up (inference on the base YAML where it belongs)

Notes

• ADS paper configs expect fully sampled CMRxRecon cine so acquisition can be simulated retrospectively. The network still only sees ACS/init plus the lines chosen by the policy.
• Install: pip install "direct-recon==2.2.0"

Contributors

• fix: ruff errors by @jonasteuwen in #315
• feat: add adaptive sampling, motion registration, MEDL, and e2e ADS configs by @georgeyiasemis in #316
• chore: bump version to 2.2.0 by @georgeyiasemis in #320

Full changelog: v2.1.1...v2.2.0