segmantR

segmantR segments microscopy images, records reproducible segmentation protocols, and exchanges masks and measurements with downstream tools. The core threshold and watershed workflows use R; optional external engines require their own runtimes and model validation.

Installation

install.packages("remotes")
remotes::install_github("CTTIR/segmantR", build_vignettes = TRUE)

See the declared R and package requirements in DESCRIPTION. Installing the R package does not install Python engines or download models.

Reproducible protocols and exchange

Besides the direct segmentation functions (sg_segment_threshold(), sg_segment_watershed(), sg_segment_propagate() and the optional Cellpose, StarDist and Mesmer wrappers), segmantR provides:

• Versioned declarative protocols (segmantR-protocol-v1): threshold.otsu.v1, threshold.adaptive.v1, threshold.triangle.v1, watershed.distance.v1, watershed.h_minima.v1, propagate.voronoi.v1, postprocess.label-cleanup.v1 and the optional stardist.2d.v1, cellpose.2d.v1, mesmer.2d.v1. Parameters are whitelisted and every run records the exact delegate call.
• A neutral interchange format (segmantR-interchange-v1): JSON manifests, exact integer label TIFF, pixel-edge GeoJSON with stable object ids, measurement tables with explicit special values, and SHA-256 inventories. Imports are staged; reviewed masks are protected.
• Training datasets and model bundles with grouped, leakage-free splits, model cards, runtime descriptions and checksums.
• QuPath programs: readable Groovy templates driven by run.json for the Script Editor and QuPath script, and a StarDist parameter adapter.

library(segmantR)
img <- sg_example_image("fluorescence_nuclei")
run <- sg_protocol_run(img, "threshold.otsu.v1", min_area = 5L, output = "run")
bundle <- sg_export_interchange(run, tempfile("bundle"), image = img)
back <- sg_import_interchange(bundle$path, expected_digest = bundle$bundle_digest)
sg_mask_status(back$mask)$status  # "staged"

The core runs without Python, QuPath or EBImage; optional backends are checked when used. Current capability status:

Contract	Status
Neutral files and inventories (I0)	supported
Protocols and run envelopes (I1)	supported
Model bundles, training datasets	supported
QuPath programs, StarDist in QuPath	planned (headless runs verified with QuPath 0.7.0 and StarDist extension 0.6.0; no partner consumer yet)
Local control service	planned

See vignette("protocols-and-interop", package = "segmantR") and system.file("interop", "INTEROP.md", package = "segmantR").

Citation

If you use this software, please cite it as:

Heller, R. (2026). segmantR: Cell segmentation of histology slides with human-in-the-loop training (Version 0.1.0) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.21889962

DOI: 10.5281/zenodo.21889962

Documentation and support

Start with the protocols and interoperability vignette or vignette("protocols-and-interop", package = "segmantR"). Use help(package = "segmantR") for the API and https://github.com/CTTIR/segmantR/issues for reproducible bug reports. The package is MIT licensed; see LICENSE.

Use of LLM tools

Where applicable, LLM tools were used for narrowly defined, non-authorial tasks: copyediting, prose refinement, Markdown/LaTeX formatting, boilerplate scaffolding (such as CI configurations and build scripts), and code refactoring. Tools included Chat AI, the KISSKI (GWDG) service, and self-hosted Mistral Small (24B, Apache-2.0) served through a local vLLM architecture on an Ubuntu Linux workstation with Intel processors and NVIDIA GPUs. The self-hosted model ran locally, with no data sent to third parties.

All scientific claims, methodological choices, analyses, interpretations and conclusions remain the authors' own. Where applicable, LLM-generated material must be reviewed and revised before inclusion, and references must be verified against their DOI, arXiv ID or ISBN.

Local software setup: vLLM with Python/PyTorch, plus R and RStudio.

Package responsibility

segmantR owns microscopy cell/nucleus segmentation, manual annotation and correction, model training, review and QuPath exchange. hyperannotatR (formerly annotatR) owns HSI annotation and evolves independently.

The offline cell drawing engine was adapted from annotatR under its MIT license (copyright 2026 Raban Heller). Its JavaScript and CSS are maintained here as sgcanvas; no annotatR or hyperannotatR runtime dependency is needed. sg_canvas(image, mask) displays cell boundaries and sg_canvas_correct() converts local pixel polygons into staged mask corrections. The app exposes rectangle, polygon, freehand, circle, edit and erase tools when htmlwidgets and magick are installed. Existing image origin/plane metadata is retained.

Histology ROI annotations

The sg_roi_* APIs represent vector annotations independently of integer instance masks and without an HSI annotation package dependency:

frame <- sg_roi_frame(c(100, 80), pixel_size = c(0.5, 0.8), pixel_unit = "um")
roi <- sg_roi_rect(10, 10, 30, 25, label = "tissue")
layer <- sg_roi_layer_add(sg_roi_layer("review"), roi)
project <- sg_roi_project(frame, layer)
mask <- sg_roi_mask(project, type = "labelled")
sg_roi_mask_stats(mask, pixel_size = frame$pixel_size)
# sg_roi_write_qupath(project, "review.geojson")
# imported <- sg_roi_read_qupath("review.geojson")

Frames declare image dimensions, calibration and optional pyramid levels; they load no pixel data. Geometry, layer editing and masks preserve image coordinates (top-left origin, x right, y down). Existing sg_mask/sg_export_qupath APIs remain available for instance-mask workflows. The ROI transfer does not include the HSI application or control service. See system.file("roi-transfer", "README.md", package = "segmantR") for provenance and limitations.

Branding

Editable R sources and regeneration instructions are in tools/branding.