MedDataGuard

MedDataGuard is a local-first dataset QA and leakage/drift auditor for medical image segmentation datasets.

This repository is intentionally code-first for portfolio/research use: core logic and controls are exposed via Python and config, without a required frontend/backend split.

What it checks

• File integrity issues (missing/corrupt images, mask/image mismatches)
• Image quality risks (blur, low contrast, resolution outliers)
• Mask quality risks (empty masks, invalid class IDs)
• Leakage (exact and near-duplicate overlap across splits)
• Drift and outliers (split-level feature shifts and PCA outliers)

Quick start

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -e .\core

Repo structure

• core/: scanner and report generation logic
• benchmark/: benchmark runner + summary generators
• demo_data/: sample data for local runs
• reports/: generated outputs

Benchmark (CholecSeg8k in archive/)

Run full benchmark:

py -3 benchmark/run_cholec_benchmark.py --archive-root archive --output-dir reports/generated

Generate portfolio summary tables:

py -3 benchmark/generate_benchmark_summary.py --input-json reports/generated/cholecseg8k_benchmark.json --output-csv reports/generated/benchmark_summary.csv --output-md reports/generated/benchmark_summary.md

Current full-run result (local, archive/):

• Samples: 8080
• Issues: 63 (image quality)
• Risk score: 0.07797 / 100
• Runtime: 640.153s
• Throughput: 12.622 samples/s
• Cross-split duplicates: exact 0, near 0

Notes

• Prototype-level heuristics; not a clinical validation system.
• Not intended for medical decision-making.