Ophthalmic Data Processing, De-identification & Quality Control Toolkit
This toolkit gathers validated, production-tested scripts written by team members for preprocessing, de-identifying, quality filtering, and structuring raw multi-center ophthalmic datasets.
1. Directory Layout & Key Modules
deidentification/: Privacy redaction & image sanitization
- audit_fundus_corner_pii.py: Optical OCR & pixel-density corner PII audit
- deidentify_trial.py: Core algorithms (fundus circular mask, frame blackout, OCR regex)
- deidentify_batch.py: High-throughput batch de-identification pipeline
- fundus_pii_trial.py: Visual inspection harness for burned-in text
- test_deidentify_trial.py: Automated regression tests for geometric masking
- crop_fundus_images.py: Automatic contour extraction & fundus border cropping
quality_control/: Multi-stage image & metadata QA
- build_quality_risk_review.py: Audits aspect ratios, corrupted headers, and outlier samples
- filter_prompt_artifacts.py: Sanitizes prompt formatting noise and deduplicates tags
- prompt_qc.py: Validates clinical prompt consistency and anatomical hierarchies
- resize_mode_qc.py: Aspect ratio and resolution distribution verification
- asrm_preprocess.py: ASRMNet-based fundus quality evaluation preprocessor
- clean_excel.py: Cross-references image files against diagnosis tables
report_parsing/: Clinical report extraction & multimodal matching
- materialize_newdata_prompts.py: Generates unified 41-column structured manifests
- attach_labels.py: Merges free-text clinical reports with image records
- batch_retrieve_reports.py: RoBERTa-SigLIP zero-shot image-to-report cross-modal retrieval
- cohorts.py: Master registry of all multi-center hospital datasets
taxonomy/: Clinical taxonomy & bilingual standardizations
- build_mixed_v5_taxonomy.py: Disease category crosswalk mapping
- build_mixed_v6_taxonomy.py: 95-disease hierarchical taxonomy builder (ICD/SNOMED aligned)
- test_build_mixed_v6_taxonomy.py: Taxonomy integrity tests
- test_taxonomy_sampling.py: Category-balanced sampling evaluation
2. Provenance & Attribution
Developed by ZJU & Multimodal Ophthalmology Lab contributors (Lisicheng, Zhanghuan, Huchengwei, Lizekun).