FreeSurfer / FastSurfer vs individual DL methods¶
Short answer¶
Keep wrapping FastSurfer / FreeSurfer for default morphometry (core, core-lesion).
Use individual engines as cited stages, utilities, and specialty modules — not as a full replacement of run_fastsurfer.sh / recon-all for SUBJECTS_DIR production.
Why wrapping is better for the default path¶
| Need | Individual tools only | FastSurfer / FS wrap |
|---|---|---|
FreeSurfer-shaped SUBJECTS_DIR (DKT, white/pial, stats) |
Fragile glue; easy to break QSIRecon/downstream | Native contract |
| Device split (GPU seg / CPU surf) | DIY | Upstream --seg_only / --surf_only |
| Versioned, cited stack | Many pins | One FastSurfer pin + modules |
| DLSurfer claim (“orchestrate, don’t replace”) | Looks like a fork | Matches claim |
DLSurfer already names the individual methods in engines.yaml (VINN, CC, CerebNet, HypVINN, recon-surf, SynthStrip, …). The planner collapses contiguous FastSurfer modules into run_fastsurfer.sh so citations stay honest without reimplementing the stack.
When individual methods are better¶
| Case | Profile / engine | Notes |
|---|---|---|
| Volume seg only | seg-quick (VINN) / seg-quick-synthseg |
No surfaces |
| Odd contrast / clinical | fs-clinical, SynthSR preprocess |
Not VINN default |
| Skull-strip / reg utils | strip, register-*, bias-deepn4 |
Standalone FS Synth* |
| Lesion fill | neurolit standalone in core-lesion |
Intentionally outside FastSurfer |
| Structures FS lacks | claustrum, SynthSeg-PC, HippUnfold, HA/BS/thal | Specialty profiles |
| Experimental CSR | surf-dl-*, SUGAR, DL-CTE |
Behind TBI±LIT gate |
What not to do¶
- Do not replace
recon-surfwith a zoo of CSR nets for production connectome handoff until the TBI±LIT gate passes. - Do not call Mauri claustrum / SynthSeg-PC as substitutes for VINN whole-brain parc.
- Do not require DKT connectome code inside DLSurfer.
Cluster practice here¶
Use Apptainer SIFs (FastSurfer 2.6, FreeSurfer, neuroLIT) via env/cluster_containers.sh so dlsurfer doctor can go READY without a host module install. Still one orchestrated stack — not a DIY VINN-only pipeline.