From “What is X-ray?” to “Let’s train a neural network”

We started with fundamentals: X-ray physics at a high level, key components (source, detector, geometry), and what makes CT different from 2D imaging. For a 15-year-old, this is far beyond typical school content—so we kept it visual, practical, and anchored in examples.
Then came the experiment.

The “Sugar vs. Salt” CT challenge

We scanned a simple sample on an FF35CT: a plastic drinking straw filled halfway with sugar and halfway with salt. In the raw projections, the difference was already visible: carbon-rich sugar vs. the higher density of NaCl. The interesting part happened after we shook the straw thoroughly to mix the grains. In 2D, the mixture became hard to separate—only small hints remained at edges.

In the reconstructed 3D volume, separation was conceptually easier… but still not trivial. 

Why thresholding wasn’t enough

A simple threshold could segment salt reasonably well, but sugar was much harder:

• sugar appeared weaker and partially blended with air and mild beam-hardening artifacts
• there were also plastic particles with similar intensity that confused the segmentation

So we moved beyond manual segmentation.

Teaching AI to do the job

Using the Segmentation Wizard in Dragonfly 3D World, we trained a neural network to classify the volume. We iterated the label strategy:

1. salt vs. sugar, with “plastic/particles/air” as a third class
2. then refined further by separating plastic and particles into additional classes

The final result is genuinely impressive—especially considering the short time frame and the intern’s age. (The visuals speak for themselves)

A practical note on speed vs. detail

To keep iteration fast and make the workflow smooth, we scanned with 2×2 binning, and then applied 2×2×2 binning in the reconstructed volume. That’s why some surfaces look a bit stepped—but it was the right tradeoff to accelerate learning and experimentation.

Main Takeaways: 

What I enjoyed most was seeing how quickly enthusiasm grows when students get meaningful tasks—real tools, real problems, real results. Pairing X-ray/CT with AI segmentation and machine learning turned the internship into something far more engaging than “watching over someone’s shoulder.”

Next up: we also worked on a second topic—3D printing + CT for deviation (as-built vs. CAD) analysis. I’ll share that in a follow-up blog.