HippUnfold local inference

The ADNI T1 task uses v1.5.3 with the bigbrain atlas. The Maguire T2 task uses the v2 development image with multihist7. The collector accepts their final space-cropT1w and space-cropT2w subfield maps, respectively. v1 omits the label-hipp entity and v2 includes it. Each hemisphere is resampled with nearest neighbour interpolation onto the model-input grid before combination. The existing left/right label map is preserved. Missing or ambiguous hemispheres stop conversion.

Before running Maguire T2, stage this file in the model bundle:

runtime-homes/hippunfold-cache/model/Dataset001_HippUnfoldMag.nnUNetTrainer__nnUNetPlans__3d_fullres.tar.gz

Zenodo record 18432751 publishes 1,138,683,578 bytes and MD5 fc2981e7b17b643b3326d3caab4912f7. Verify both before inference. The repository manifest additionally records SHA-256 ee4e28d246cb1acb86df6b99be50c66b9d762356385643092877a9723a552de7. The runner checks that the archive exists and is nonempty; it does not compute the checksum on every run. For a different host location, pass --model-archive /path/to/the/archive.tar.gz to run_hippunfold.py.

Keep the pre-staged template/CITI168, template/upenn, and atlases/tpl-multihist7 directories under that cache, and the prebuilt runtime-homes/hippunfold-v2-overlay-dir beside it. The cache, archive, and v2 runtime overlay are mounted read-only. All derived outputs and Snakemake work files go under the job work directory.

The v2 image's download_nnunet_model rule uses storage(model_dict["url"]). Setting the cache directory alone does not remove its HTTP dependency. Only the HTTP storage plugin is bundled, so a filesystem URL is insufficient. hippunfold_offline.smk uses documented Snakemake module rule exclusions to import the bundled processing rules without model, template, or atlas download/copy rules. The staged resources become ordinary external inputs; the model archive is no longer a temporary output that Snakemake can delete. An incomplete resource cache fails with a local missing-input error. No container or bundled workflow files are modified.

This adapter targets the bundled v2 workflow and its rule names. Check it again when updating the image. CPU regression tests cover collection, resampling, missing outputs, and command construction. They do not establish GPU inference success or clinical readiness.