Lung nodule detection from CT scans

Reading chest CT scans for lung cancer screening means working through hundreds of image slices per patient, where an early catch changes outcomes and a missed nodule can cost a life. Enzzero built a deep learning system that automates the first pass: find the nodules, score how suspicious each one looks, and show the evidence behind every call.

What we built

• A 3D convolutional neural network with Squeeze-and-Excitation attention that reads CT volumes the way radiologists do: in three dimensions, not slice by slice
• One multi-task model with three heads: nodule detection, malignancy scoring on the 1 to 5 radiologist consensus scale, and spatial localization
• Grad-CAM saliency maps so a reviewer can see exactly which tissue drove each prediction, plus automatic Lung-RADS categorization for reporting
• Training on LIDC-IDRI, the reference public dataset: 1,018 chest CT scans annotated by up to four expert radiologists, prepared into 5,250 volumetric patches across 875 patients

The numbers

On the held-out test set (88 patients the model never saw during training), detection reached an AUC of 0.997 with 97% patch-level accuracy. Malignancy predictions landed within about 0.9 points on the 1 to 5 clinical scale, inside the 1 to 2 point range where expert radiologists disagree with each other on the same nodule.

What we say out loud

This is a research system, not a cleared medical device. When we audited it at whole-scan level on an independent external dataset, performance dropped below what deployment would demand, so the same report publishes those audit numbers and the retraining they triggered. That is the standard we hold every build to: measure honestly, publish the audit, improve what the audit exposes.

Why it matters

Most of our work saves teams hours. This build shows what the same discipline looks like when the stakes are clinical: reference datasets, adversarial evaluation, and predictions a reviewer can interrogate instead of a black box.

This system was researched, trained, and audited end to end by Ahura, one of Enzzero's engineers.

See how Enzzero approaches custom AI development.