histoseg 0.1.7

...

HistoSeg

HistoSeg is a Python toolkit for spatial transcriptomics segmentation / geometry extraction.

The current focus is Pattern1 isoline (0.5) contour generation from cell clusters (e.g., 10x Xenium GraphClust output):

• Pick a set of “target clusters” (Pattern1)
• Fit a KNN regressor to estimate P(target) over space
• Smooth the probability field
• Extract a contour (isoline) at level = 0.5
• Save contour vertices and a quick preview plot

---

Quick links

• Documentation: https://histoseg.readthedocs.io/en/latest/
• Source code: https://github.com/hutaobo/HistoSeg
• Issue tracker: https://github.com/hutaobo/HistoSeg/issues

⚠️ License note

This project is distributed under the PolyForm Noncommercial 1.0.0 license. Academic and other noncommercial use is permitted. Any commercial use requires a separate commercial license from SPATHO AB. See LICENSE for the full terms.

---

Installation

Install from PyPI (recommended)

pip install -U histoseg

Install from source (for development)

git clone https://github.com/hutaobo/HistoSeg.git
cd HistoSeg
pip install -U pip
pip install -e .

Dependencies

The Pattern1 isoline workflow uses:

• numpy, pandas
• scipy
• scikit-learn
• matplotlib
• a Parquet engine (pyarrow is recommended)

If you run into missing imports, install them explicitly:

pip install -U numpy pandas pyarrow scipy scikit-learn matplotlib

Optional:

• Hugging Face downloader: pip install -U huggingface_hub

---

Tutorial: Pattern1 isoline (0.5)

What you need (inputs)

The isoline workflow expects the following files:

1. clusters.csv

   • Typically from GraphClust: analysis/clustering/gene_expression_graphclust/clusters.csv
   • Must contain columns: Barcode, Cluster

2. cells.parquet

   • A cell-level table with spatial coordinates (x/y-like columns)
   • Must contain at least:
     • coordinate columns (e.g. x/y or x_centroid/y_centroid)
     • an id column that can be aligned with clusters.csv:Barcode (the code tries several common column names)

3. tissue_boundary.csv (optional but recommended if you enable synthetic background)

   • Must contain columns x,y or X,Y

What you get (outputs)

By default, the pipeline writes into out_dir:

• params.json — all parameters + inferred join columns
• pattern1_isoline_<level>_<i>.npy — contour vertices (Nx2 arrays)
• pattern1_isoline_<level>.png — quick preview plot

---

Quickstart

One-liner (from a Hugging Face dataset repo)

This follows the example notebook in examples/contour_generation_pattern1_from_hf.ipynb.

# pip install -U histoseg
# pip install -U huggingface_hub pandas pyarrow numpy scipy scikit-learn matplotlib

from histoseg import run_pattern1_isoline_from_hf

PATTERN1 = (10, 23, 19, 27, 14, 20, 25, 26)

result = run_pattern1_isoline_from_hf(
    repo_id="hutaobo/output-XETG00082_C105",
    revision="main",  # or a commit hash for strict reproducibility
    out_dir="outputs/pattern1_isoline0p5_from_graphclust",
    pattern1_clusters=PATTERN1,

    # Defaults are intentionally exposed for tuning:
    grid_n=1200,
    knn_k=30,
    smooth_sigma=5.0,
    min_cells_inside=10,
)

print("Outputs folder:", result.out_dir)
print("Preview image:", result.preview_png)
print("Contours:", len(result.contours))

Run on local files

from histoseg import Pattern1IsolineConfig, run_pattern1_isoline

PATTERN1 = (10, 23, 19, 27, 14, 20, 25, 26)

cfg = Pattern1IsolineConfig(
    clusters_csv="/path/to/analysis/clustering/gene_expression_graphclust/clusters.csv",
    cells_parquet="/path/to/cells.parquet",
    tissue_boundary_csv="/path/to/tissue_boundary.csv",
    out_dir="outputs/pattern1_isoline0p5",
    pattern1_clusters=PATTERN1,

    # Optional tuning:
    grid_n=1200,
    knn_k=30,
    smooth_sigma=5.0,
    min_cells_inside=10,
)

result = run_pattern1_isoline(cfg)
print(result)

---

How it works (workflow overview)

flowchart TD
  A["clusters.csv<br/>Barcode/Cluster"] --> C["Align barcodes<br/>with cells.parquet"]
  B["cells.parquet<br/>x/y + id-like column"] --> C
  C --> D["Select target clusters<br/>(Pattern1)"]
  D --> E["Sample background points<br/>(other cells)"]
  F["tissue_boundary.csv"] --> G["Generate synthetic background<br/>(optional)"]
  G --> E
  D --> H["KNN regression<br/>predict P(target)"]
  E --> H
  H --> I["Predict on mesh grid"]
  I --> J["Gaussian smoothing"]
  J --> K["Mask by tissue<br/>(nearest-cell threshold)"]
  K --> L["Extract isoline<br/>level = 0.5"]
  L --> M["Filter loops<br/>min_cells_inside"]
  M --> N["Save params.json<br/>+ contours .npy<br/>+ preview .png"]

---

Troubleshooting & tuning

If no contour is found, try:

• Decrease min_cells_inside (e.g. 10 → 3)
• Increase smooth_sigma (e.g. 5 → 8)
• Increase knn_k (e.g. 30 → 50)
• Reduce grid_n to speed up (note: grid_n=1200 can be heavy)

---

API reference (high-level)

Pattern1 isoline

• Pattern1IsolineConfig
  Dataclass holding all parameters and input paths.

• run_pattern1_isoline(cfg) -> Pattern1IsolineResult
  Runs the full pipeline on local files.

• run_pattern1_isoline_from_hf(repo_id, revision="main", ...) -> Pattern1IsolineResult
  Convenience wrapper that downloads required files from a Hugging Face dataset repo and then runs the pipeline.

Hugging Face I/O helpers

• download_xenium_outs(repo_id, revision="main", clusters_relpath=..., cache_dir=None)
  Downloads cells.parquet, tissue_boundary.csv, and the specified clusters.csv from a dataset repo.

SFPlot utilities (legacy / optional)

This repository contains a small subset of SFPlot-style utilities and re-exports:

• compute_cophenetic_distances_from_df(df, ...)
• plot_cophenetic_heatmap(matrix, ...)

---

GUI (experimental)

A GUI entry point is configured as:

histoseg-gui

Notes:

• The current GUI code path is still in flux and may require extra dependencies (e.g., Pillow) and/or an external sfplot installation.
• For production workflows, prefer the Python API shown above.

---

Contributing

Issues and pull requests are welcome.

When reporting a bug, please include:

• OS + Python version
• histoseg version
• Minimal reproducible code (or a small input subset)
• Expected vs. actual behavior

---

License

This project is distributed under the PolyForm Noncommercial 1.0.0 license. Noncommercial use (including academic research) is permitted. Any commercial use requires a separate commercial license from SPATHO AB. See LICENSE for details.