26,000 Lung Scans Get Labeled So AI Can Spot Cancer Earlier

A team of medical imaging researchers has finished labeling a huge pile of lung scans. The scans come from the National Lung Screening Trial, a landmark U.S. study that followed over 26,000 people at risk of lung cancer. In this new work, they added detailed digital tags to those images — marking where each organ, blood vessel, and suspicious nodule appears. Then they posted everything in a free public cancer imaging database anyone can use.

Why does labeling matter so much? Think of it like flash cards for a student. An AI system learns to spot tumors by studying thousands of examples where an expert has already circled the answer. Without those circles, the AI is guessing. Radiologists are in short supply and their time is expensive, so most scans sit unlabeled. This project turns more than 26,000 previously silent scans into teaching material — like handing a medical student a fully annotated textbook instead of a blank one.

The numbers are big. One dataset covers 26,194 patients and automatically outlines organs in each scan using a tool called TotalSegmentator (software that draws organ boundaries for you), plus measurements of texture and shape for every region. Two smaller sets, covering 581 and 601 patients, hold expert-reviewed annotations of actual lung nodules. All of it is stored in a standard medical format called DICOM, so it plugs directly into hospital software and research tools. Basically, they removed the busywork that kept these images out of reach.

The catch: this is research infrastructure, not a medical product. The scans come from a screening population, so they don't represent every kind of patient or disease, and machine-generated outlines can contain errors. Nobody should expect an app to diagnose them tomorrow. Still, this is the unglamorous plumbing that makes better screening tools possible — and lung cancer is the deadliest cancer worldwide, largely because it's usually found late.