hmnshudhmn24/noisy-medical-document-images-ocr

🏥 Noisy Medical Document Images (OCR)

1,000 noisy, synthetic medical document images with structured JSON ground truth — built for Document AI, LayoutLM fine-tuning, and clinical NLP research.

---

🧭 Overview

This dataset provides 1,000 high-resolution images of two healthcare document types, each degraded with realistic scanning artifacts to simulate real-world OCR conditions:

Category	Count	Description
🧾 Hospital Bills	500	Itemized statements with CPT codes, insurance adjustments, barcodes, financial summaries
📄 Discharge Summaries	500	Clinical narratives with HPI, hospital course, labs, medications, physician signatures

Generated with ReportLab and degraded using a multi-stage OpenCV noise pipeline — fully synthetic, 100% HIPAA-safe.

---

✅ Why This Dataset

• Realistic noise: perspective warps, folds, stains, punch holes, blur, JPEG artifacts
• Zero real patient data — fully synthetic
• Rich, structured JSON ground truth for every field
• Two document classes for classification + OCR tasks

---

📊 Ground Truth Schema

Column	Type	Description
filename	string	Image file name
document_type	string	bill or discharge_summary
json_data	JSON string	Structured metadata (patient, diagnosis, ICD-10, procedures, meds, financials)

---

🔬 Noise Pipeline

Stage	Effect	Probability
Perspective Warp	Skewed scan angle	100%
Mail Folds	Crease lines	70%
Uneven Lighting	Shadow gradient	100%
Coffee Stains / Punch Holes	Physical artifacts	~15% each
Blur + JPEG Compression	Quality degradation	60%

---

⚡ Quick Start

import pandas as pd import json from PIL import Image

bills_df = pd.read_csv("medical_bills_ground_truth.csv") row = bills_df.iloc[0] data = json.loads(row["json_data"])

print(data["patient"]["name"], data["diagnosis"]) img = Image.open(f"bills/{row['filename']}") img.show()

---

💡 Use Cases

• OCR training & evaluation (Tesseract, EasyOCR, PaddleOCR)
• LayoutLM / DiT fine-tuning
• Document classification (bill vs discharge summary)
• Clinical NER (patient, diagnosis, ICD-10, meds)
• Table extraction from CPT-coded bills
• VLM benchmarking (GPT-4V, Gemini Vision, LLaVA)