Choi TMLI geometry and label review

The task stays disabled. The axis variant and numeric mapping of the 18 regions remain unconfirmed. CPU geometry tests and local review sheets do not establish clinical or manuscript readiness.

Voxel-space geometry finding

Choi et al. (2024) report non-contrast CT, mean pixel size 1.113 × 1.113 mm, and 5 mm slices. The released model metadata record training sizes of 219–275 × 512 × 512 in array z/y/x order, 1 × 1 × 1 mm spacing for all 20 cases, and transpose_forward=[0,1,2]. The full-resolution stage uses 1 mm. These observations support dummy-spacing, voxel-space training. They do not identify the authors' storage-axis directions. The mean pixel size also does not establish each training case's exact physical spacing.

The previous runner supplied real CT spacing. nnU-Net therefore resampled a 1.5 mm slice stack to 1 mm: a 735-slice scan became about 1102 slices, compared with about 221 slices at the paper's 5 mm scale. This is approximately a fivefold cranio-caudal scale mismatch and a 1.113-fold in-plane scale mismatch. The mismatch is verified; whether it explains the incoherent predictions requires the manager's variant and fold comparisons.

Input and inverse geometry contract

run_choi_tmli.py accepts a 3D CT in HU with explicit millimetre units. It uses all eight voxel-face corners to cover the physical field of view, including feet-first, prone and oblique acquisitions:

1. Resample HU linearly at 1.113 × 1.113 × 5 mm in physical LPS axes. Outside HU is −1024. Centre each target grid on the physical bounding box. Each in-plane dimension is at least 512; pad smaller fields and enlarge a dimension when its field exceeds 512 × 1.113 = 569.856 mm. Preserve spacing and never crop. The slice count is the ceiling of physical superior–inferior coverage / 5, allowing 0.0001 voxel tolerance for NIfTI float32 header rounding.
2. Specify three flip bits using --axis-variant or CHOI_AXIS_VARIANT. The bits refer to storage x/y/z relative to identity LPS directions, as read by SimpleITK from an HFS DICOM stack. 000 means L/P/S positive directions; 100 reverses x, 010 reverses y, 001 reverses z. All eight combinations are available. These are physical directions, independent of input storage.
3. Add 1024 once after HU resampling. Write the same array with 1 mm spacing, zero origin and identity LPS direction. Its NIfTI RAS affine is diag(-1,-1,1,1). nnU-Net's SimpleITK reader thus sees the training header; it may crop background but performs no spacing change at full resolution.
4. Restore the prediction's true target geometry from geometry.json, then resample integer labels with nearest neighbour onto the original input grid. Export the source shape, affine and millimetre units for converter checks.

The default is unconfirmed; inference fails before creating a workspace unless an explicit variant is supplied. The following selects an evaluation variant, not a confirmed default:

python run_choi_tmli.py --input case.nii.gz --model RELEASE_TRAINER_FOLDER \
  --work NEW_PRIVATE_WORK_DIRECTORY --output choi.nii.gz \
  --axis-variant 000 --folds 0

Do not supply an already shifted CT. Omitting --folds uses all five folds. The normal nnU-Net v1 predictor averages fold probabilities before argmax; restored discrete label maps are never averaged. Test-time mirroring is disabled. Checkpoints are read through links; mutable metadata and outputs use the work directory. No confirmed default is selected by the evaluation script.

Generate anatomical evidence

Use the original CT in HU, the restored numeric prediction, and separate binary TotalSegmentator total masks on the same CT grid. Dependencies are nibabel, numpy, scipy and Pillow.

python scripts/map-choi-labels.py --ct case.nii.gz \
  --prediction choi.nii.gz --total total_masks --output review

Add --fold-agreement fold-agreement.json when the manager script has generated pairwise fold Dice. Agreement describes the individual folds, including on an ensemble sheet; it is not a comparison against a clinical reference contour. Both-absent labels have undefined Dice and one-absent labels have zero Dice.

review.html starts with an 18-label summary: volume, component count, largest component share, its side and vertebral span, the top two candidate regions with scores, and fold agreement when available. Per-label axial/coronal panels follow. Raw measurements are inside collapsed details. choi-review-template.csv has columns label, proposed region, reviewer region, confidence, comment; reviewer fields start blank. Completing a CSV does not confirm the label map.

The largest 6-neighbour component determines the displayed side, span and candidate-region evidence. A separate span covers the union of components individually at least 1 mL. Full-label extents remain in anatomy.json so fragments remain inspectable. Vertebral spans use nearest centroids and intersecting physical z ranges, not vertebral boundary assignments. The report retains the three largest component centroids and volumes. Widely separated substantial components, competing rules and duplicate proposals lower confidence. Scores are heuristic evidence strengths, not calibrated probabilities.

All measurements use the full NIfTI RAS affine. Patient left is negative RAS x, positive LPS x. Side uses the local spinal midline with a 3 mm central band and an 80% side threshold. Midline extrapolation is flagged. Review panels place patient right at image left; true axial and coronal planes also handle obliquity. The panels sample the largest component; inspect the full volume for extent.

Input files are identified by SHA-256. Measured grids must match after lossless axis permutations/flips. A mask may inherit millimetre units only when its unspecified-unit grid matches the explicitly millimetre CT. No registration is inferred. TotalSegmentator lacks separate obturator fossa, femoral head, carina, and internal/external iliac vessel masks; these proxies cannot settle pelvic or mediastinal subdivisions. Missing labels require further folds or cases.

Manager checks and mapping

The private manager script first compares fold 0 across eight variants using largest-component share, with absent labels scored zero. It then evaluates all five folds of the best two, ranks by mean inter-fold Dice followed by coherence, and runs the best variant's five-fold probability ensemble. Runtime and sampled process-group GPU memory, per-label volumes/components/spans and all pairwise fold Dice are recorded. Exploratory coherence thresholds do not select a clinically validated orientation or establish label identity.

Result of the first full screen (2026-10-15, one internal whole-body CT): no variant met the coherence screen. The best mean inter-fold Dice was 0.291 (variant 000) and 0.250 (variant 110), and the five-fold ensemble also failed. Folds took 2.7–11 min with the corrected geometry. Until a variant is coherent, the numeric labels are not sent for clinical review and the task stays disabled.

After manager geometry review, a clinician must resolve all 18 numeric labels and record review evidence. The HTML can export a complete, attributed mapping JSON; labels/choi-tmli.json accepts no partial or duplicate assignments. Preflight and dispatch reject unconfirmed mappings. Fresh inference, converter, and orientation checks remain necessary. Keep patient data and reviewer identities in private storage. Catalog enablement remains a separate manager decision after the axis contract and label mapping are resolved.

Sources checked 2026-10-15: PLOS ONE paper, Zenodo release.