FiberMargin

FiberMargin performs training-free repair of categorical masks and spatial labels. It is a post-processing method: upstream segmentation, clustering, or annotation remains unchanged. A deterministic atlas of rotated Hilbert paths transports leave-self-out class evidence from both directions using normalized path length only. Its class score is the two-sided enclosure (sqrt(left) + sqrt(right))^2. FiberMargin changes a label only when its full-atlas margin clears a conservative chart-disagreement barrier; locally isolated points receive extra protection at that decision.

The package exposes one production refiner with four arguments. It does not need expression values, a graph, training labels, or user-selected tuning parameters.

Installation

install.packages("remotes")
remotes::install_github("tkcaccia/fibermargin", build_vignettes = TRUE)

Basic Use

library(fibermargin)

sim <- simulate_gradient_regions(
  n = 50000,
  minority = 0.25,
  samples = 2,
  density_profile = "strong"
)

refined <- refine_spatial_labels(
  xy = sim$xy,
  labels = sim$labels,
  samples = sim$samples
)

c(
  input = mean(sim$labels == sim$truth),
  refined = mean(refined == sim$truth)
)

Only xy and labels are required. samples defines independent coordinate systems: coordinates may overlap completely between specimens, but neither coordinates nor labels cross a specimen boundary. Constant coordinate axes are removed separately within each specimen. A specimen with columns (x, y, z) and constant z is therefore exactly equivalent to (x, y), while genuinely three-dimensional specimens use all three coordinates.

workers is the total native CPU budget for the call, not a budget per specimen:

refined <- refine_spatial_labels(xy, labels, samples, workers = 4)

The same budget is reused as specimens are processed, so there is no nested process pool or worker multiplication. Native parallelism is available on macOS, Linux, and Windows. Changing workers does not change labels or inferential diagnostics; only the workers bookkeeping attribute records the effective budget.

On an 8-core Apple M3, the release benchmark for 100,000 genuinely 3D locations in four specimens took a median 0.518 seconds with one worker and 0.183 seconds with four workers (2.83x speedup; three measured repetitions per budget). Validation, specimen indexing, axis reduction, coordinate normalization, and native refinement are included; simulation generation is excluded. The script and raw timings are under benchmarks/benchmark_release_workers.R and benchmarks/results/release_0.1.0_worker_benchmark/.

Genuine 3D coordinates

volume <- simulate_volumetric_domains(
  n = 12000,
  shape = "folded_layers",
  samples = 2,
  seed = 42
)

volume_refined <- refine_spatial_labels(
  volume$xy, volume$labels, volume$samples, workers = 4
)
attr(volume_refined, "dimensions_used")

Independent specimens with overlapping coordinates

xy_one <- sim$xy[sim$samples == levels(sim$samples)[1], , drop = FALSE]
xy <- rbind(xy_one, xy_one)
samples <- rep(c("section_1", "section_2"), each = nrow(xy_one))
labels <- factor(rep(sim$labels[sim$samples == levels(sim$samples)[1]], 2))

joint <- refine_spatial_labels(xy, labels, samples, workers = 4)
attr(joint, "sample_sizes")

Clean a Categorical Mask

For a 2D pixel mask or 3D voxel array, coordinates are generated automatically. Missing cells remain missing. No source image, intensity, model logit, or clean-mask training collection is used.

truth <- matrix(rep(c("background", "region"), each = 5000), nrow = 100)
noisy <- corrupt_categorical_mask(
  truth, mechanism = "boundary", rate = 0.15, seed = 4
)
cleaned <- clean_categorical_mask(noisy, workers = 1)

evaluate_mask_cleaning(truth, noisy, cleaned)

The evaluator reports multiclass mean IoU, one-grid-step Boundary IoU, rare-class IoU, correction recall, and damage. Its accuracy obeys the exact identity gain = repaired_fraction - damaged_fraction.

Diagnostics

The result is an ordinary factor with the input levels. Pointwise diagnostics are attached without expanding the public interface:

attr(refined, "candidate")
attr(refined, "margin_score")
attr(refined, "required")
attr(refined, "repair_margin")
attr(refined, "atlas_dispersion")
attr(refined, "isolation")
attr(refined, "changed")

margin_score is an uncalibrated rival-versus-observed evidence contrast, not a probability or expected loss. repair_margin >= 0 is the selective change criterion for a candidate that differs from the observed label. atlas_dispersion is the empirical standard deviation of the deterministic chart contrasts divided by the square root of the atlas size. It is a descriptive admission scale, not a calibrated probability or confidence interval. isolation is the multiclass local-gap protection factor; it is one for the binary ballot.

The following named summaries are also attached:

attr(refined, "workers")
attr(refined, "dimensions_used")
attr(refined, "labels_changed")
attr(refined, "changed_fraction")
attr(refined, "classes_before")
attr(refined, "classes_after")
attr(refined, "removed_classes")
attr(refined, "sample_sizes")

All summaries except workers are reported per specimen; class summaries are named lists. FiberMargin does not force the output to retain every observed class. removed_classes makes any disappearance explicit.

Evaluation

The package provides planar, gradient, complex-shape, and volumetric simulators, plus one evaluator used by every benchmark:

evaluate_spatial_refinement(
  truth = sim$truth,
  initial = sim$labels,
  refined = refined,
  boundary = sim$boundary,
  regions = sim$area,
  sparse = sim$sparse,
  method = "FiberMargin"
)

correction_recall is the fraction of initially wrong labels repaired. damage_rate is the fraction of initially correct labels made wrong. Worst-class, sparse-region, and boundary accuracy prevent aggregate accuracy from hiding local failures.

Compact DLPFC and CRC Visium HD coordinate/annotation benchmarks are bundled. CRC denotes colorectal cancer. DLPFC is redistributed under the spatialLIBD terms. The CRC derivative is CC BY 4.0 and contains only coordinates, 19 author-derived region labels, and stored corruption recipes; expression counts and tissue imagery are excluded. MERFISH inputs remain external. See inst/extdata/REAL_DATA_LICENSES.md for sources, attribution, and change notices.

crc <- load_spatial_benchmark(
  "crc", "CRC_random_25_r1", seed = 1040001L
)
mean(crc$labels != crc$truth)

Current Validation

The C++ implementation is evaluated on predefined CRC, DLPFC, and MERFISH corruption protocols. On the direct-comparator matrices, it obtains 0.8589 accuracy and 0.7806 ARI across 60 CRC corruptions, within 0.0015 and 0.0021 of multiscale mode. Across 45 DLPFC corruptions it obtains 0.8942 accuracy and 0.8220 ARI. Across 45 MERFISH corruptions it obtains 0.8905 accuracy and 0.8004 ARI, exceeding the strongest fixed comparator by 0.0062 accuracy and 0.0116 ARI. These are controlled label-recovery results, not claims that the biological annotations are naturally wrong.

The protocol definitions, package-build metadata, per-case ledgers, and matched plots are under benchmarks/results/fibermargin_atlas912_full_roster_v1/; the source-matched FiberMargin confirmation is in benchmarks/results/fibermargin_atlas912_full_confirmation_v1/. Additional simulation and mask panels are retained in the repository and are re-evaluated only when the production source changes.

See vignette("fibermargin", package = "fibermargin") for the workflow, vignette("benchmarks", package = "fibermargin") for benchmark design, and vignette("fibermargin_reference", package = "fibermargin") for the package reference API.