naazimsnh02/medocr-vision-dataset

MedOCR-Vision Dataset

A curated OCR dataset for training PaddleOCR models on medical documents while maintaining general OCR capabilities through domain-balanced training.

Dataset Description

Dataset Summary

MedOCR-Vision is a carefully curated dataset combining medical and general documents to train robust OCR models. The dataset prevents catastrophic forgetting by maintaining a balanced distribution between medical (59.4%) and general (40.6%) domains.

• Total Samples: 2,462
• Domain Balance: 59.4% Medical / 40.6% General
• Splits: 80% Train / 10% Validation / 10% Test
• Languages: English
• License: MIT
• Author: Syed Naazim Hussain
• Created: December 2025

Supported Tasks

• Optical Character Recognition (OCR): Extract text from medical and general documents
• Document Understanding: Comprehend structured and unstructured document layouts
• Medical Text Extraction: Specialized extraction from medical prescriptions, lab reports, and forms

Languages

The dataset contains text in English, including:

• Medical terminology
• Patient information
• Test results and values
• Business documents (invoices, receipts)
• Dates, numbers, and structured data

Dataset Structure

Data Instances

Each instance in the dataset contains an image and its corresponding text annotation:

{
    'image': <PIL.Image.Image>,
    'text': 'Patient Name: John Doe\nAge: 45\nGender: Male\nDiagnosis: ...'
}

Data Fields

• image: A PIL Image object containing the document image (RGB format)
• text: A string containing the ground truth text extracted from the image

Data Splits

Split	Samples	Percentage	Size (MB)
Train	1,969	80.0%	408 MB
Validation	246	10.0%	53 MB
Test	247	10.0%	48 MB
Total	2,462	100%	509 MB

Dataset Composition

Source Datasets

Source	Samples	Domain	Type	Origin
Medical Prescriptions	1,000	Medical	Handwritten	HuggingFace
OMR Scanned Documents	36	Medical	Scanned Forms	HuggingFace
Medical Lab Reports	426	Medical	Printed Reports	Kaggle
Invoices & Receipts	1,000	General	Business Docs	HuggingFace

Domain Distribution

• Medical Domain (1,462 samples - 59.4%):

  • Handwritten prescriptions
  • Scanned OMR forms
  • Printed medical laboratory reports
  • Medical terminology and patient data

• General Domain (1,000 samples - 40.6%):

  • Business invoices
  • Receipts
  • Structured tables and forms
  • Dates, numbers, and financial data

Text Characteristics

• Short (200-400 characters): Prescriptions, Invoices
• Medium (400-1,000 characters): OMR Documents
• Long (1,000-5,000 characters): Medical Lab Reports

Dataset Creation

Source Data

Medical Prescriptions

• Source: chinmays18/medical-prescription-dataset
• Platform: HuggingFace
• Processing: Direct extraction from JSON annotations
• Samples: 1,000

OMR Scanned Documents

• Source: saurabh1896/OMR-scanned-documents
• Platform: HuggingFace
• Processing: LLM-based text extraction using Nebius API
• Samples: 36

Medical Lab Reports

• Source: dikshaasinghhh/bajaj
• Platform: Kaggle
• Processing: Parallel LLM-based extraction using Nebius and Novita APIs
• Samples: 426

Invoices & Receipts

• Source: mychen76/invoices-and-receipts_ocr_v1
• Platform: HuggingFace
• Processing: Direct extraction from existing annotations, randomly sampled
• Samples: 1,000 (sampled from 2,043)

Annotations

Annotations were created through multiple methods:

• Direct Extraction: For datasets with existing ground truth annotations
• LLM-based Extraction: Using Nebius and Novita APIs for image-to-text conversion
• Quality Verification: Manual verification and deduplication
• Format Standardization: Consistent formatting across all sources

Data Processing Pipeline

1. Collection: Downloaded source datasets from HuggingFace and Kaggle
2. Processing: Applied appropriate extraction methods per dataset
3. Cleaning: Removed duplicates and verified annotations
4. Balancing: Sampled invoices/receipts to achieve 60/40 medical/general ratio
5. Splitting: Created 80/10/10 train/validation/test splits with stratification
6. Validation: Verified 100% image-annotation matching

Usage

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("naazimsnh02/medocr-vision-dataset")

# Access individual splits
train_data = dataset['train']
val_data = dataset['validation']
test_data = dataset['test']

# View dataset structure
print(dataset)

Basic Usage Example

from datasets import load_dataset
from PIL import Image

# Load dataset
dataset = load_dataset("naazimsnh02/medocr-vision-dataset")

# Iterate through training data
for sample in dataset['train']:
    image = sample['image']  # PIL Image
    text = sample['text']    # Ground truth text
    
    # Your processing code here
    print(f"Text length: {len(text)} characters")

Use with PaddleOCR Training

from datasets import load_dataset
from paddleocr import PaddleOCR

# Load dataset
dataset = load_dataset("naazimsnh02/medocr-vision-dataset")

# Initialize PaddleOCR
ocr = PaddleOCR(use_angle_cls=True, lang='en')

# Training loop
for batch in dataset['train']:
    images = batch['image']
    labels = batch['text']
    
    # Your PaddleOCR training code here

Data Visualization

from datasets import load_dataset
import matplotlib.pyplot as plt

dataset = load_dataset("naazimsnh02/medocr-vision-dataset")

# Display a sample
sample = dataset['train'][0]
plt.imshow(sample['image'])
plt.title(f"Text: {sample['text'][:100]}...")
plt.axis('off')
plt.show()

Considerations for Using the Data

Intended Use

This dataset is intended for:

• Training OCR models for medical documents
• Research in document understanding
• Developing healthcare AI applications
• Preventing catastrophic forgetting in domain-specific models

Out-of-Scope Use

This dataset should NOT be used for:

• Clinical decision making without human oversight
• Replacing professional medical transcription
• Processing real patient data without proper authorization
• Any use that violates patient privacy or medical ethics

Privacy and Ethics

• All medical data is from publicly available datasets
• No real patient identifiable information is included
• Data has been processed to ensure privacy compliance
• Users should follow appropriate data protection regulations

Additional Information

Dataset Curators

Syed Naazim Hussain

• GitHub: naazimsnh02
• HuggingFace: naazimsnh02

Licensing Information

This dataset is released under the MIT License.

MIT License

Copyright (c) 2025 Syed Naazim Hussain

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

Citation Information

If you use this dataset in your research, please cite:

@dataset{medocr_vision_2025,
  title={MedOCR-Vision: Medical OCR Dataset with Balanced Domain Training},
  author={Syed Naazim Hussain},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/naazimsnh02/medocr-vision-dataset}
}

Acknowledgments

• Thanks to all original dataset creators for making their data publicly available
• Nebius and Novita for providing LLM API access for text extraction
• PaddleOCR team for the excellent OCR framework
• HuggingFace for hosting and dataset infrastructure

---

Version: 1.0
Last Updated: December 2025