Logo Lanfrica

Software0408/nigeria_cxr_benchmark_phase1

Domain:

healthcare

Record type:

dataset
Creator:
Sof
Host:
Phase 1 foundations for Nigerian/African CXR benchmark – evaluation pipeline only # Nigerian Chest X-ray AI Benchmark — Phase 1 > **Evaluation-only benchmark** for assessing commercial AI chest X-ray models against a Nigerian/African patient population. > Solo research project at Federal Medical Centre (FMC) Ebute-Metta, Lagos. > Target venue: MICCAI / MIDL 2027. --- ## Table of Contents 1. Project Overview 2. Ethics & Data Governance 3. Compute Environment 4. Label System 5. NLP Pipeline — Phase A (Weeks 6–8) - 5.1 Rule-Based Extractor (Week 6) - 5.2 BioClinicalBERT Extractor (Week 7) - 5.3 Synthetic Augmentation & Ensemble (Week 8) 6. Results Summary 7. Full Dataset Statistics (5,517 Reports) 8. Directory Structure 9. Installation & Setup 10. Running the Pipeline 11. Testing 12. CI/CD 13. Project Roadmap --- ## 1. Project Overview This project constructs the first publicly documented Nigerian chest X-ray benchmark for evaluating AI diagnostic models. Phase 1 covers NLP-based label extraction from free-text radiology reports; subsequent phases will cover model evaluation, cross-dataset experiments, and interpretability. **Scope (Phase 1):** | Component | Detail | |---|---| | Dataset | 5,517 anonymised Nigerian CXR radiology reports | | Models to evaluate (Phase D) | DenseNet-121, CheXNet, EfficientNet-B4 | | Comparison datasets (Phase E) | CheXpert, MIMIC-CXR, VinDr-CXR | | Interpretability (Phase F) | Grad-CAM | | Domain adaptation | Deferred to Phase 2 | --- ## 2. Ethics & Data Governance - **Ethics approval:** FMC Ebute-Metta Research Ethics Committee — approved **March 3, 2026** - **Anonymisation:** Patient identifiers are replaced with deterministic hashed IDs using HMAC-SHA256. The anonymisation salt (`CXR_ANON_SECRET`) is stored as an encrypted environment variable under PI control and is **never committed to this repository**. - **Storage:** Raw patient data lives in a VeraCrypt-encrypted volume (`workspace/`) tracked with DVC. Only derived, non-identifiable artefacts are committed to Git. --- ## 3. Compute Environment | I …

Languages

Licenses