Perforated OCR

Inspiration

Doctors’ handwritten prescriptions are often hard to read, leading to errors and delays in healthcare. We wanted to see if Perforated AI’s dendritic optimization could push OCR performance beyond what standard models achieve on such noisy, real-world data.

What it does

Perforated OCR enhances a Transformer-based OCR model (TrOCR) with dendritic learning to better recognize illegible handwritten medical prescriptions, improving accuracy and learning efficiency on low-quality images.

How we built it

We started with TrOCR as a baseline, integrated Perforated AI’s dendritic optimization, trained on handwritten prescription samples, and evaluated how dendrites reduce residual errors compared to the standard model.

Challenges we ran into

• Highly variable and messy handwriting
• Low-resolution and noisy scans
• Integrating and tuning dendritic components with a large Transformer model

Accomplishments that we're proud of

• Successfully applying dendritic optimization to a real-world OCR task
• Demonstrating measurable improvement over the baseline
• Delivering a clean, reproducible hackathon-ready implementation

What we learned

We learned how dendritic learning can enhance feature selectivity and robustness, especially in difficult perception tasks where traditional deep networks still struggle.

What's next for Perforated OCR

• Train on larger and more diverse medical datasets
• Extend to multilingual handwritten prescriptions
• Explore deployment in clinical decision-support systems to reduce reading errors and improve patient safety

Built With

• ai
• dendrite
• machine-learning
• ml
• perforated
• python
• pytorch