PneuNet: eighteen months of chest X-rays, three competition rounds

Heilbronn is done. We got the special prize for prevention and rehabilitation from the German Social Accident Insurance (DGUV), 1,000 euros. Not a national win. For something that started as a school research paper, I’ll happily take it.

It started with a number we found early on: diagnosing pneumonia from chest X-rays by eye is right about 76 percent of the time, and rural areas don’t have enough radiologists. Nico, Dustin and I wanted to know whether a neural network could help doctors there. That became our extended research project at the specialist maths and science school in Erfurt, starting in eleventh grade. We worked with a university hospital on it, which mostly took persistence.

I didn’t start with PyTorch. I started with NumPy: my own network, backpropagation written by hand. Slow, but afterwards I knew what the frameworks actually do. Then the systematic comparison. ResNet, DenseNet, EfficientNet, ConvNeXt and VGG19 on four public datasets with up to 30,000 images, plus lung segmentation with an attention U-Net and class activation maps so you can see where the model is looking. In the end it all ran inside a DICOM viewer.

The viewer is the part you can actually touch, and for me it’s the real product. DICOM is the format hospitals store X-rays in, image and patient data in one file. A model that only runs in a notebook doesn’t help any doctor. So the model has to go where doctors already look. The viewer shows the files and a patient list, with the X-ray in the middle. It runs the models over the image and shows the result right inside it: pneumonia or healthy, the lung mask from the segmentation, and the class activation map as a heatmap. The heatmap shows which regions counted most for the decision. If it lights up next to the lung instead of inside it, something is wrong, no matter how confident the model sounds. That’s exactly what it’s for. There’s a full version in C++ with Qt and a test version in Python. The demo on this page rebuilds the Python version with public sample images and a lens for the heatmap. It doesn’t make a diagnosis.