Troubleshooting HyperASHS¶

Installation¶

hrashs check reports problems with PyTorch or CUDA. HyperASHS requires PyTorch 2.5.x. Install the PyTorch build that matches your CUDA version, as described under Advanced setup, and then reinstall HyperASHS.

Segmentation is very slow. HyperASHS is probably running on the CPU. Check that nvidia-smi lists your GPU and that hrashs check finds CUDA. You can force the GPU with --device cuda.

Models cannot be downloaded from Hugging Face (SSL errors). If you are behind a corporate or institutional proxy, add the -k option to the hrashs command to disable SSL verification (see SSL certificate errors).

Results¶

The segmentation is in the wrong place, or the QC images show misaligned images. Check the registration QC images in the qc folder. If the T1w and T2w scans are not aligned, register them manually and pass the transform with --reg-t1t2, or use --reg-t1t2-identity if they are already aligned (see Registration options).

The segmentation looks blocky. You are probably looking at the native-space segmentation (*_tse_native_seg_*). Use the hyper-resolution segmentation (*_tse_patch_hyperres_seg_*) for measurements (see Outputs).

The T1-only run prints a warning. T1-only mode is only recommended for models that do not label hippocampal subfields, such as PMC_T1_Atlas.

Getting help¶

Ask on the ashs-users mailing list, or open an issue on GitHub. Please include the hrashs command you ran, the output of hrashs check, and the QC images if relevant.

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