TBI ± LIT — surface DL gate report¶
Date: 2026-10-23–24
TopoFit gate: passed (tbi_lit_topofit) — machine + Chamfer soft QC
V2C gate: open (tbi_lit_surfaces — literature-best Ready full CSR)
SUGAR / DL-CTE: separate open gates (sphere / thickness-only)
Piriform: production stays JLF until SynthSeg-PC eval (synthseg_pc_vs_jlf)
Literature picks (one per stage)¶
| Stage | Chosen | Why |
|---|---|---|
| Full CSR (Phase 3) | V2C-Flow | Best Ready joint WM+pial + template IDs (Bongratz MedIA 2024); over CortexODE/PialNN |
| sphere.reg alt | SUGAR | Multi-dataset / speed (Ren 2023) > S3Reg |
| Thickness w/o mesh | DL-CTE family | Fast biomarker path; never for connectome |
| Piriform production | JLF | Best deployed IDEAS method until PC clears Hammers bar |
TopoFit — machine + Chamfer QC¶
Score: work/phase3_topofit/topofit_tbi_lit_score.json
FreeView: work/phase3_topofit/freeview_topofit_qc.sh
| Metric | with-LIT lh/rh | no-LIT |
|---|---|---|
*.white.topofit |
yes / yes | yes / yes |
| mean Chamfer vs recon-surf white | ~2.2–2.7 mm | similar |
| hausdorff_95 | ~10–14 mm | elevated (unequal topology) |
machine_ok=true, chamfer_soft_ok=true (≤4 mm). Visual QC: run FreeView script (red=TopoFit+LIT, yellow=noLIT, green=recon-surf).
V2C-Flow — install path¶
- Clone:
Documents/others/Vox2Cortex(pretrainedV2C-Flow-S-ADNI) - Wrapper:
env/bin/dlsurfer-v2c - Job:
sbatch hpc/sbatch_phase3_v2c_tbi_lit.shafterconda env create -f requirements.yml(+ pytorch3d) or their Docker wheel underVox2Cortex/docker/ - Gate stays open until ±LIT inference succeeds and meshes are scored
Baseline recon-surf (EG)¶
Standalone LIT (core-lesion) validated on EG vs native LIT — see prior COMPARISON_EG.md. GPU confirmatory job 191960 COMPLETED (work/eg_gpu_core_lesion, recon-surf.done).