Surface metric and IBSI validation

Synthetic controls test physical grids, area-weighted surface distances, empty masks, and contour agreement. Comparison with the pinned DeepMind implementation covers 96 random mask pairs and 384 tolerance-specific comparisons. HD95 agrees exactly; maximum absolute differences are 1.78 × 10⁻¹⁵ for pooled mean distance and 4.44 × 10⁻¹⁶ for surface Dice. Upstream comparison. Synthetic sphere measurements.

The IBSI receipt records 735 feature outputs that match the retained public benchmark bit for bit, and 300 preprocessing diagnostic comparisons within the published tolerances. A, B, D and E each meet tolerance for 140/141 computed feature comparisons; C meets tolerance for 109/141. PyRadiomics kurtosis and fixed-bin-width differences remain in the results. Reference coverage is 141/346 pairs for A/B and 141/210 for C/D/E.

The synthetic default extraction contains 1,223 radiomic features and 37 diagnostic entries. Its serialized fingerprint is 4174115f0c5c677b17b56ea525606cbda57639c999035618ada18b69bcc08fd7. Tests cover separate morphology and intensity masks, non-unit labels, quality-control counts, parameter identity and configured feature classes.

Reproduce the package-to-benchmark comparison from the repository root:

OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 \
  "${RTPIPELINE_RADIOMICS_PYTHON:-python}" -I -B \
  tests/validate_f6_ibsi.py --output docs/validation/f6_ibsi

The radiomics environment needs NumPy 1.x and PyRadiomics. Public inputs are restored with the pinned retrieval procedure in validation/ibsi/README.md. The receipt verifies 26 public input hashes and 22 retained raw extraction hashes.

These measurements cover public phantoms and synthetic masks. Native CT extraction supports this preprocessing option; the Conda helper and isolated robustness factories do not. Finite-resolution sphere bias and feature-formula differences are quantified in the feature documentation. The clinical ingestion and publication workflow is outside this validation.