Exploring AI in Healthcare: Image Classification

Kian van Holst, a first-year Applied Data Science and Artificial Intelligence student, built an image classification system for healthcare - and chose to make it harder for himself by going for three classes rather than a binary split.

The system

It classifies chest X-rays into healthy lungs, lungs with bacterial infection, and lungs with viral infection, trained on data sourced from Kaggle.

Building it

Kian set baseline metrics before building anything, then concentrated on preprocessing. The model uses Convolutional Neural Networks - “neural networks capable of processing 2D images” - with filter layers that pick out the differences that matter and tie them to outcomes. It is supervised learning on labelled examples, written in Python.

The first version reached 83% on the three-class problem. Applying transfer learning, reusing pre-trained models, cut training time substantially.

What the confusion matrix showed

Telling bacterial from viral infection turned out to be much harder than telling healthy from unhealthy. The binary healthy/unhealthy split reached 94% - and as Kian notes, that is the distinction that matters most for patient care.

Where it could go

Beyond radiology itself, he proposed an educational app for radiology students. Adding explainable AI methods such as GradCAM heatmaps, which show where the model is looking, would let students learn to spot lung conditions while seeing how the model reached its own conclusion.