NA-MIC Project Weeks

Back to Projects List

Recon-all correction script based on manual subcortical segmentation files

Key Investigators

• Jarrett Rushmore (Center for Morphometric Analysis, Massachusetts General Hospital, USA)
• Benoît Verreman (ETS, Canada)
• Sylvain Bouix (ETS, Canada)

Project Description

White and pial surfaces and parcellation (aparc+aseg) from recon-all pipeline (FreeSurfer 7.4.1) need quite a few manual corrections. At the same time, 100 cases from HCP-YA dataset were manually segmented using HOA2 atlas. The goal of the project is to leverage the latter to get better results with recon-all.

Objective

1. Make the pipeline more straightforward, by using only one edited input image (subcortical HOA segmentation)
2. Visually check the different scenario results of the script, and solve following region issues: hippocampus/amygdala, temporal lobe
3. Train an nnUNet model on segmenting white and gray matter based on manually edited ribbon

Approach and Plan

1. Create a new script to merge HOA and FS aseg.presurf.mgz file (merge_hoa_into_aseg.py)
2. Modify previous script (ribbon_edit_script.sh) to adapt to new input
3. Use edited ribbon of HOA to train an nnUNet model

Progress and Next Steps

1. merge_hoa_into_aseg.py works properly
2. Final result is good: surfaces and parcellation are HOA compatible (minimal corrections needed)
3. nnUNet trained model didn’t segment properly WM/GM
   • add HOA subcortical labels to help segmentation
   • compare with other models (like Unest)

Illustrations

https://github.com/user-attachments/assets/8df68d33-37ac-4b2e-b94d-73d2219f620b