LUNA25: 3D Deep Learning on Lung CT — Work in Progress

Problem

The LUNA25 challenge asks whether lung nodules in CT scans are malignant. Before any model can compete, there's an unglamorous engineering problem: the dataset is a multi-part archive of large 3D volumes that has to be downloaded, reassembled, matched to labels, and fed to a GPU efficiently.

What's built so far

• Parallel dataset assembly — multi-threaded download and reassembly of the Zenodo archive parts
• Label matching — SeriesInstanceUID-keyed joins organizing scans into labeled splits
• 3D volume loading — SimpleITK pipelines producing 128³ voxel inputs
• Baseline 3D CNN — a first Keras convolutional architecture wired end to end

Honest status

This is an experiment in progress, and no performance claims are made for it: the current training runs use a small sampled subset, so their metrics are not meaningful generalization estimates and are deliberately not quoted here. It's included in the portfolio because the data-engineering half — getting 3D medical imaging from a public archive into a trainable pipeline — is real, finished work, and representative of the practice's approach: infrastructure first, honest metrics second.