Agentic Spatial Pathologist

Agentic workflows for spatial pathology.

spatho is the Python package and CLI for Agentic Spatial Pathologist. It wraps the lower-level histoseg engine with workflow configuration, organ packs, artifact manifests, H&E overlays, structure review, and report generation.

The platform framing is:

stGPT learns reusable contour/region morpho-molecular representations; spatho plans, validates, and turns them into auditable spatial pathology evidence.

Legacy standalone surface: the canonical product-layer implementation is being integrated into ASTRO under app/src/xenium_ai_discovery/pathology_app/. This repository remains the compatibility, packaging, and deployment-oriented shell for spatho.

Two Parallel AI Backends

spatho now supports two parallel AI review paths. The paid OpenAI API route remains available and unchanged, while the local pathology-ai route adds a self-hosted option for private or cluster deployments.

Backend	How it runs	Best for	Key settings
openai	Calls the paid OpenAI API with your OPENAI_API_KEY.	Fast setup, managed models, lightweight local machine.	pathology_review_backend="openai"
pathology_ai_api	Calls a local HTTP service backed by vLLM, embeddings, reranking, and Qdrant.	PDC/HPC, private data, cost control, local model operations.	pathology_review_backend="pathology_ai_api" and pathology_ai_api_base_url

The two paths are intentionally independent: enabling the local service does not remove or disable the OpenAI backend.

Foundation Evidence Workbench

spatho is the agentic spatial pathology workbench layer. It can consume precomputed stGPT artifacts without importing stgpt, or call a local stGPT install when stgpt_backend="local_stgpt" is configured. stGPT evidence is guarded before biological review: missing required artifacts make spatho doctor not ready, fatal QC blocks a run, and warning-only QC is shown as cautionary model-derived evidence.

Workflow fields default to disabled:

{
  "stgpt_enabled": true,
  "stgpt_backend": "precomputed_artifacts",
  "stgpt_artifact_dir": "/path/to/stgpt/spatho_export",
  "stgpt_min_cell_coverage": 0.95,
  "stgpt_require_qc_pass": true
}

Expected stGPT artifacts are cell_embeddings.parquet, structure_embedding_summary.csv or structure_summary.parquet, and qc_report.json. The workbench writes stgpt_evidence_summary.csv/json, updates the artifact manifest, and inserts a report section that labels the evidence as model-derived rather than measured expression.

Quick Start

Install from PyPI

python -m pip install -U spatho

For local development:

git clone https://github.com/hutaobo/Agentic-Spatial-Pathologist.git
cd Agentic-Spatial-Pathologist
python -m pip install -U pip
python -m pip install -e .[dev]

If you are actively developing against a local histoseg checkout, install that editable copy first:

python -m pip install -e ../HistoSeg

Path A: paid OpenAI API

Use this path when you want the simplest managed-model setup.

export OPENAI_API_KEY=sk-...
spatho init-workflow \
  --organ breast \
  --case-name breast_case_01 \
  --dataset-root /path/to/Xenium_outs \
  --base-pipeline-config /path/to/project/configs/breast_case_01.json \
  --output /path/to/workflows/breast_case_01_openai.json

spatho run --config /path/to/workflows/breast_case_01_openai.json

The generated workflow can keep:

{
  "pathology_review_backend": "openai"
}

Disable OpenAI and force heuristic mode when needed:

spatho run --config /path/to/workflow.json --heuristic-only

Path B: local pathology-ai service

Use this path when you want pathology review to call a self-hosted service instead of the paid OpenAI API.

{
  "pathology_review_backend": "pathology_ai_api",
  "pathology_ai_api_base_url": "http://localhost:8000"
}

For PDC/Dardel deployment, see docs/PDC_LOCAL_PATHOLOGY_AI.md. The local stack is:

• pathology-ai: lightweight HTTP orchestration from this repo
• vllm: OpenAI-compatible local LLM endpoint
• embedder: TEI-compatible Python embedding service for BAAI/bge-m3
• reranker: TEI-compatible Python reranking service for BAAI/bge-reranker-v2-m3
• qdrant: local vector storage

Default local model configuration:

LLM_BASE_URL=http://127.0.0.1:8001/v1
LLM_MODEL=openai/gpt-oss-120b
EMBED_MODEL=BAAI/bge-m3
RERANK_MODEL=BAAI/bge-reranker-v2-m3
VECTOR_DB=qdrant
DEFAULT_TOP_K=6
STRICT_JSON=true

Common CLI Tasks

Check an environment and workflow config:

spatho doctor --config /path/to/workflow.json

List built-in organ packs:

spatho list-organ-packs

Export the workflow JSON schema:

spatho config-schema --output /path/to/workflow.schema.json

Build or refresh an artifact manifest:

spatho build-manifest --config /path/to/workflow.json

Write Xenium RNA+protein + H&E alignment fixtures:

spatho write-xenium-alignment-fixtures \
  --output-dir /path/to/output/pipeline/validation \
  --segmentation-source ranger_protein_assisted

This writes a Xenium RNA+protein alignment note, a fixture manifest, and transform cases covering identity, um -> pixel, translation, axis order, and composed polygon export.

Python Usage

from spatho import run_evidence_workbench, run_workflow

result = run_workflow("/path/to/workflows/breast_case_01_openai.json")
print(result["pathology_report_html"])

workbench_result = run_evidence_workbench("/path/to/workflows/breast_case_01_openai.json")

Generate a starter config from Python:

from spatho import init_workflow

result = init_workflow(
    "/path/to/workflows/breast_case_01_openai.json",
    organ="breast",
    case_name="breast_case_01",
    dataset_root="/path/to/Xenium_outs",
    base_pipeline_config="/path/to/project/configs/breast_case_01.json",
)
print(result["workflow_config"])

What a Workflow Produces

A typical full run produces:

• cluster evidence bundles
• OpenAI, local pathology-ai, or heuristic cluster annotations
• dendrogram-guided structure assignments
• clustermap and H&E overlay artifacts
• structure-level pathology reviews
• case-level HTML report
• machine-readable artifact manifest

Organ Packs

spatho ships with built-in organ packs that define the annotation taxonomy, default study context, workflow parameter defaults, and expected artifact contract.

Built-in packs:

• lung
• breast

These packs live in src/spatho/organ_packs.

Config Contract

Workflow JSON files are backed by a formal schema exported from the package. For Xenium RNA+protein workflows, the config template records:

• dataset_modality = xenium_rna_protein
• canonical_space = physical_um
• export_space = xenium_explorer_pixel
• xenium_pixel_size_um
• segmentation_source

See docs/XENIUM_RNA_PROTEIN_ALIGNMENT.md for the rationale and polygon-level analysis model.

Repository Layout

• src/spatho: public-facing Python package and CLI
• src/pathology_ai_service: local pathology AI HTTP service
• deploy/pathology_ai: Docker Compose and PDC Slurm/Apptainer deployment assets
• docs/PDC_LOCAL_PATHOLOGY_AI.md: local/PDC pathology-ai deployment guide
• docs/PYPI_RELEASE.md: PyPI publishing checklist
• examples/workflows: public-safe starter workflow templates
• main.py: older Gradio/Serve deployment surface kept for compatibility

Relationship to HistoSeg and ASTRO

Current implementation model:

• histoseg executes the geometry, segmentation, and workflow internals
• spatho wraps and presents the workflow as a product-facing package

Target implementation model:

• histoseg remains the geometry/segmentation engine
• spatho owns workflow UX, organ packs, public docs, reports, and deployment surfaces
• the canonical integrated product implementation continues to move into ASTRO

Publishing

This repo includes a PyPI publishing workflow based on GitHub Actions Trusted Publishing. See docs/PYPI_RELEASE.md for setup and release steps.

License

This project is intended for noncommercial research use unless separately licensed. Before public release or commercial use, review the license text and commercial boundary together with the underlying histoseg dependency.