Building an AI Assistant for Chest X-rays

Why I Built This

After nearly a decade working in radiologic imaging, I’ve seen firsthand how overburdened radiologists can become. My goal with this project was to explore how machine learning—specifically convolutional neural networks (CNNs)—could help ease some of that burden.

How It Works

Using the NIH Chest X-ray dataset, I trained a CNN using TensorFlow and Keras to classify images as either normal or abnormal. The idea isn’t to replace radiologists, but to offer a simple triage tool that flags potentially abnormal cases so they can be prioritized.

The model doesn’t diagnose—it highlights. It’s a small step toward workflow automation in medical imaging, and one that could meaningfully reduce delays in patient care.

Challenges & Takeaways

Working with real medical images presented unique challenges around preprocessing, memory limits, and ensuring HIPAA compliance. I learned a lot about batch processing, augmentation, and building interfaces that keep human review at the center.

What’s Next

I chose not to make this model directly usable through the site for obvious ethical reasons. Still, the architecture could be adapted for future models in other imaging domains like CT or MRI. For now, it’s a proof-of-concept I'm proud of—blending healthcare with code in a meaningful way.