JoselinMaria/medical-schedule-ocr

🏥 Medical Schedule OCR Pipeline

Local OCR extraction pipeline for scanned medical schedule PDFs.
Extracts begin_time, end_time, patient_name, dob (+ room, surgeon, anes, age/sex, minutes) per row.

96.9% field accuracy on synthetic benchmark — 100% on time fields, 93.8% on names and dates.

Quick Start

# Install
pip install easyocr pypdfium2 opencv-python-headless numpy pandas Pillow

# Run on your PDFs
python extract_schedules.py --input-dir C:\ml\nt-pdfs --output-dir ./output

# With GPU + higher DPI
python extract_schedules.py -i ./pdfs -o ./out --dpi 400 --gpu --verbose

What It Does

PDF pages → grayscale → deskew → denoise → CLAHE → EasyOCR → header detection
→ row grouping → column assignment → regex normalisation → validation → CSV + JSON

Files

File	Purpose
extract_schedules.py	Main pipeline — the only file you need to run
create_test_pdf.py	Generate synthetic test PDFs with known ground truth
validate.py	Compare extraction results against ground truth
RECOMMENDATIONS.md	Engine selection, tuning guide, known limitations

Output

• schedule_extracted.csv — one row per schedule entry with confidence score
• schedule_extracted.json — same data with per-field confidence breakdown
• extraction_summary.txt — human-readable report
• debug/ — preprocessed page images, annotated bounding boxes, raw OCR text

Expected Column Layout

ROOM | BEGIN | END | MIN | SURGEON | ANES | PATIENT'S NAME | AGE/SEX | DOB

The header is auto-detected by keyword matching. If your PDFs use different headers, edit HEADER_KW in the script.

CLI Options

--input-dir, -i    Folder with scanned PDF files (required)
--output-dir, -o   Output folder (default: ./output)
--dpi              Render DPI (default: 300, try 400 for small text)
--row-threshold    Y-pixel tolerance for row grouping (default: 25)
--gpu              Use GPU for EasyOCR
--no-debug         Skip debug output
--verbose, -v      Verbose logging

Requirements

• Python 3.10+
• No system packages needed (no poppler, no tesseract)
• CPU works fine; GPU speeds up OCR ~3-5x