hvf_extraction_script

hvf_extraction_script extracts and structures data from Humphrey Visual Field (HVF) single field analysis report images using OCR (Tesseract) and image processing (OpenCV) to parse reports generated by HFA2 and HFA3 devices for extraction of metadata, value plot data, and percentile plot data.

Key Features:

• Automated data extraction: Extracts and organizes metadata, value plot data, and percentile plot data from HVF single field analysis report images.
• OCR engine: Employs Tesseract for optical character recognition of textual report elements.
• Image processing: Uses OpenCV for image preprocessing and region parsing of HVF report images.
• Supported devices/formats: Targets HVF single field analysis reports produced by HFA2 and HFA3 devices.
• Validation dataset: Validated on 90 HVF reports across three report layouts comprising 1,530 metadata fields, 15,536 value plot data points, and 10,210 percentile data points.
• Benchmarking against standards: Performance was compared to manual extraction by four human operators with DICOM reference data as the gold standard.
• Performance metrics: Processing time 4.9–8.9 seconds per report (over 40× faster than 6.5–19 minutes for human extractors) and reported error rates: metadata 1.2%–3.5% (human 0.2%–9.2%); value plot data v1 0.9%, v2 <0.01%, v3 0.15% (human 0.8%–9.2%); percentile data v1/v2 0%, v3 0.06% (human 0.06%–12.2%).

Scientific Applications:

• High-throughput visual field data processing: Enables large-scale extraction and structuring of HVF datasets for quantitative analysis.
• Ophthalmology research: Supports extraction of metadata and sensitivity/value maps for studies of visual field outcomes and disease progression.
• Clinical data validation: Facilitates comparison of extracted report data with DICOM references for quality assurance and benchmarking of extraction methods.

Methodology:

Uses Tesseract OCR and OpenCV image processing to parse HVF single field analysis report images from HFA2/HFA3; validation comprised analysis of 90 reports across three layouts, comparison to manual extraction by four operators, and use of DICOM as the reference standard.