Models and Atlases¶

The pre-trained HyperASHS models and the ASHS Classic atlas packages are hosted on Hugging Face, under the upennpatchlab account.

HyperASHS models and templates¶

You don't need to download these yourself

hrashs run downloads the model you select, and the template it uses, automatically the first time you use them, and keeps them in a local cache. Cloud HyperASHS needs no download at all. The repositories are listed here for reference, for example to read a model's description, check its version, or download models in advance for a computer without internet access.

All HyperASHS repositories are at https://huggingface.co/upennpatchlab.

Segmentation models. Each repository holds a trained model together with its description (inputs, labels and citations), which you can also view with hrashs desc.

Model ID	Hugging Face repository	Input MRI
`PMC_Atlas`	upennpatchlab/hyperresashs_model_pmc_uni_v1	T2w + T1w
`PMC_T1_Atlas`	upennpatchlab/hyperresashs_model_pmc_uni_t1_v1	T1w only
`HSG_Body_Atlas`	upennpatchlab/hyperresashs_model_hsg_body_v1	T2w + T1w

Templates. Models use a template to locate and crop the medial temporal lobe (MTL) region in whole-brain MRI. Each model names the template it needs, and hrashs downloads it along with the model.

Template	Hugging Face repository
Aging 3T (used by the current models)	upennpatchlab/hyperresashs_template_aging3t_v1
Whole brain (for tasks that work on the whole brain)	upennpatchlab/hyperresashs_template_wholebrain_v1

Model directory. upennpatchlab/hyperresashs_atlas_directory lists the models that hrashs list shows.

If you train your own model, the hrashs upload command can publish it to Hugging Face.

ASHS Classic atlas packages¶

ASHS Classic atlas packages can be downloaded from the upennpatchlab/ashs_atlases dataset on Hugging Face, as well as from NITRC. Unlike HyperASHS models, these must be downloaded and unpacked by hand before running ASHS locally (see Installation). You don't need them for Cloud ASHS.

Each package is a .tgz archive of 2–5 GB, named after the atlas and its release date (YYYYMMDD). The dataset includes:

Package	Atlas
`ashs_atlas_upennpmc_20170810.tgz`, `ashs_atlas_upennpmc_20210803.tgz`	ASHS-PMC: Penn Memory Center, 3T T2-weighted MRI
`ashs_atlas_upennpmc_t1_20180720.tgz`, `ashs_atlas_upennpmc_t1_20230717.tgz`	ASHS-PMC-T1: Penn Memory Center, 3T T1-weighted MRI
`ashs_atlas_upennpmc_t1ext_20240617.tgz`	Penn Memory Center, T1-weighted MRI, extended version
`ashs_atlas_abc_prisma_20200406.tgz`	ABC atlas, 3T (Prisma)
`ashs_atlas_abc_prisma_7t_20200622.tgz`	ABC atlas, 7T
`ashs_atlas_magdeburg_7t_20180416.tgz`	Magdeburg 7T atlas
`ashs_atlas_neuropresage_lc_20230331.tgz`	NeuroPresage atlas

See the README of each package for its segmentation protocol and citation.

To download a single package from the command line, you can use the Hugging Face CLI:

pip install -U huggingface_hub
hf download upennpatchlab/ashs_atlases ashs_atlas_upennpmc_20210803.tgz \
    --repo-type dataset --local-dir ./atlases