Skip to content

Quick Start

HyperASHS Quick Start

This section is organized as a series of questions to help you get started with HyperASHS.

1. Pick the HyperASHS model that is best for your data

Most users will run HyperASHS with one of the pre-trained models that we provide. For this to work, your MRI scans must be sufficiently similar to the scans used to train the model. Currently, the following models are available:

PMC_Atlas: Penn Memory Center T2w + T1w 3T MRI model

Use this model if:

  • Your MRI protocol includes
    • T2-weighted MRI of approximately 0.4 × 0.4 × 2.0 mm³ resolution, with slices orthogonal to the main axis of the hippocampus
    • T1-weighted MRI of approximately 1.0 × 1.0 × 1.0 mm³ resolution
  • You wish to segment hippocampal subfields and extrahippocampal MTL subregions
  • You are studying older adults (studies of aging and early neurodegenerative diseases)

T2-weighted input image

Input T2-weighted MRI, upsampled to the resolution of the segmentation.

PMC_Atlas segmentation

HyperASHS segmentation with the PMC_Atlas model.

PMC_T1_Atlas: Penn Memory Center T1w-only 3T MRI model

Use this model if:

  • Your MRI protocol includes only whole-brain T1-weighted MRI (no high-resolution T2)
  • You wish to segment extrahippocampal MTL subregions (ERC, BA35, BA36 and PHC), the amygdala and the hippocampus (anterior/posterior, but not hippocampal subfields)
  • You are studying older adults (studies of aging and early neurodegenerative diseases)

This model uses the same segmentation protocol as PMC_Atlas, so measures of the MTL cortex are consistent between the two models.

T1-weighted input image

Input T1-weighted MRI, upsampled to the resolution of the segmentation.

PMC_T1_Atlas segmentation

HyperASHS segmentation with the PMC_T1_Atlas model.

HSG_Body_Atlas: Hippocampal Subfields Group body model

Use this model if:

  • You have T2w + T1w MRI as for PMC_Atlas, and
  • You want hippocampal subfields in the hippocampal body, following the harmonized protocol of the Hippocampal Subfields Group (HSG)

Train a custom model

If none of the models above suit your needs, you can train your own. You will need:

  • MRI data from 20–30 representative subjects with manual segmentations
  • A Linux computer with an NVIDIA GPU

See Training a Model.

2. Verify that your MRI data is suited for HyperASHS

Check that your data meet the requirements for the model you choose:

3. Choose how to run HyperASHS

You can run HyperASHS in two ways:

  • By sending your data to a cloud-based server that runs HyperASHS for you (the easy way)
  • By installing HyperASHS on your own computer (best for large studies and for training your own model). A Linux computer with an NVIDIA GPU is strongly recommended. A segmentation takes about 3 minutes on a GPU, but over an hour on a CPU.

To run HyperASHS locally:

  1. Install HyperASHS:

    pip install hyperresashs
hrashs check

  2. Run the segmentation:

    # T2w + T1w MRI
hrashs run -a PMC_Atlas -g t1.nii.gz -f t2.nii.gz -I sub01 -w ./sub01

# T1w MRI only
hrashs run -a PMC_T1_Atlas -g t1.nii.gz -I sub01 -w ./sub01

Using classic ASHS? See the equivalent page for ASHS Classic.