Deciphering Doctor's Handwriting with Deep Learning

Explore a deep learning approach to deciphering doctors' handwriting in this 51-minute Devoxx conference talk. Delve into the development of a data-processing pipeline designed to support human data entry of medical information from handwritten documents. Learn about the challenges of digitizing handwritten medical records and how machine learning can improve accuracy and efficiency. Discover the end-to-end process, from data mining and anonymization to image processing and deep learning model construction. Examine real-life examples and statistics demonstrating the model's effectiveness in reducing illegible handwritings by 30%. Gain insights into the application of AI modeling for doctor's handwriting across various use cases, including death certificates and consultation certificates. Follow the speaker's journey through data preprocessing, model training, confidence level calculations, and integration of natural language processing techniques. Understand how this innovative approach compares to traditional OCR methods and addresses challenges such as language detection, certificate type identification, and complex handwriting patterns.

Intro
Business goal
Case 2: the Doctor's Consultation Certificate
Case 1: the Death Certificate
Data pipeline
Mining
Original image
Line segmentation
Baseline
Training set definition
Data quality
Data partitioning: split into text lines
Deep-Learning Model
Ensemble modelling
Model Training Performance
Model accuracy
Prediction confidence levels
Confidence plot
Examples target population
Examples false positives
Statistics false positives
Examples low confidence
Example: type A
Language detection
Certificate type detection
Preprocessing
Training data
Nomenclature code
Date of Consultation
Comparison OCR - Neural Network
The Grid (5)
Some grid examples
Approach
Examples: 2 lines
Training on histogram features
Results number of lines prediction
Reading the lines
Examples: high confidence false predictions
Grid summary
Second challenge: different patterns
Fourth challenge: rotations
Fifth challenge: superposition and readability
Step 1: find the stamp
Intersection over Union
Summary stamp reading
Application Integration
Conclusion