AngioStress v0.1 Real-Data-First Benchmark

AngioStress is a deterministic benchmark package for stress-testing frozen endovascular perception models on digital subtraction angiography and endovascular segmentation surfaces.

The benchmark core is real data:

• DIAS sequence segmentation: 20 labeled test sequences, 115 evaluated frames per model, 345 prediction rows across three frozen models.
• CathAction human segmentation: 5,225 audited nonempty image-mask pairs, 15,675 prediction rows across three frozen models.

The TopCoW-derived synthetic projection fixture is auxiliary. It is used for controlled stressor and regression checks, not as the core benchmark surface.

Frozen Model Panel

• SAM ViT-B (sam_vit_b)
• SAM ViT-L (sam_vit_l)
• MedSAM ViT-B (medsam_vit_b)

All model entries are treated as frozen, off-the-shelf evaluation targets. This repository does not train or fine-tune a new backbone.

Main Benchmark Files

• Contract: benchmark/contracts/angiostress_v0_1_real_data.json
• Contract validator: benchmark/validate_contract.py
• DIAS benchmark runner: benchmark/run_dias_contract_full_test.py
• CathAction full-tier benchmark runner: benchmark/run_cathaction_full_tier.py
• CathAction quick subset materializer: benchmark/run_cathaction_contract_subset.py
• Release audit: benchmark/run_release_audit.py
• Public release staging/audit: benchmark/stage_public_release.py
• DIAS outputs: experiments/main/run-dias-contract-full-test-v0/outputs/
• CathAction full-tier outputs: experiments/main/run-cathaction-real-scaleup-v0/outputs/full_nonempty_5225/

Public Artifacts

• GitHub code and lightweight benchmark package: https://github.com/SeldingerMed/angiostress-benchmark
• Hugging Face derived benchmark artifacts: https://huggingface.co/datasets/txmedai/angiostress-benchmark

Install The Package

The benchmark utilities are installable with the standard Python packaging workflow:

python3 -m pip install .

The package currently uses only the Python standard library at runtime. Development/test helpers are declared as an optional extra:

python3 -m pip install '.[dev]'

Console entry points:

• angiostress-validate-contract
• angiostress-run-dias-contract
• angiostress-run-cathaction-subset
• angiostress-run-cathaction-full-tier
• angiostress-run-release-audit
• angiostress-stage-public-release

Validate The Package

Run the release audit from the repository root:

python3 benchmark/run_release_audit.py

Expected output:

• experiments/main/run-real-benchmark-release-refresh-v0/outputs/release_audit.json
• experiments/main/run-real-benchmark-release-refresh-v0/outputs/release_manifest.json
• experiments/main/run-real-benchmark-release-refresh-v0/outputs/metrics_summary.json
• experiments/main/run-real-benchmark-release-refresh-v0/outputs/validation_summary.md

The audit checks that DIAS and CathAction are both present as core real-data surfaces, that copied derived prediction artifacts exist, that metrics are finite, and that private manuscript/LaTeX paths are excluded from the release manifest.

Stage the public upload package:

python3 benchmark/stage_public_release.py

Data-Source Boundary

DIAS and CathAction source datasets should be obtained from their original sources and licenses. This package records benchmark outputs, derived predictions/overlays, manifests, metrics, and provenance; it should not be treated as a redistribution of raw clinical source data.

Claim Boundary

AngioStress v0.1 is a benchmark artifact and measurement package. The public package exposes real DIAS and CathAction evaluation surfaces, frozen-model metrics, manifests, and derived prediction artifacts. It should not be presented as clinical validation, model improvement, or proof of synthetic-to-real transfer.