Standalone pythomics feature extraction

Extract per-object lumen and epithelium measurements and block-summed density maps from an integer label-map image. This folder is self-contained and does not import the repository's model/inference code.

Install

From this directory, install the listed dependencies into your environment:

python -m pip install -r requirements.txt

Image readers are provided by scikit-image. Install glymur (python -m pip install glymur) to enable multi-threaded JPEG2000 decoding; otherwise JPEG2000 files use the scikit-image reader. Object-level tables are Parquet, density arrays are compressed NPZ, and MAT output is optional.

CLI

Run from the repository root:

python -m pythomics_standalone path/to/labels.png --output-dir output --save-mat

To regenerate a MAT file using an existing label map and object tables:

python -m pythomics_standalone path/to/labels.png \
  --lumen-parquet output/labels_lumen.parquet \
  --epithelium-parquet output/labels_epithelium.parquet \
  --output-dir output

The default label values are lumen=1, epithelium=2, stroma=4, epithelial-cells=5, stromal-cells=6. Override with --lumen-label, --epithelium-label, --stroma-label, --epithelial-cells-label, and --stromal-cells-label. --min-area sets the minimum connected component size in source pixels (default 2048); --block-size HEIGHT WIDTH sets density aggregation blocks (default 20 20).

Epithelium regions include both epithelium and epithelial-cell pixels. Stroma density includes stromal-cell pixels. Connected components are labeled independently in each tissue mask. Tables contain source-pixel area, circularity (roundness), and for epithelium average skeleton thickness and epithelial-cell fraction. Density NPZ arrays are block sums, with edge blocks padded with zeros. MAT density maps are block sums; per-object maps at density resolution are pixel-area-weighted means of the feature values for included objects in each block, excluding background and regions filtered out by --min-area. --save-full-mat writes those per-object maps at source-pixel resolution instead.

Outputs are <input-stem>_lumen.parquet, <input-stem>_epithelium.parquet, and <input-stem>_densities.npz; with --save-mat, also <input-stem>_features.mat at density (block) resolution, or full resolution with --save-full-mat. To build a MAT file from an existing label map and tables, pass --lumen-parquet and --epithelium-parquet. -j/--jobs sets worker threads for JPEG2000 decoding and large-image mask, density, component-labeling, and feature-map calculations (default: up to 12 CPUs); specify a higher value to use more.