AI Model Combining Imaging and Blood Tests Improves Lung Cancer Diagnosis

Lung cancer kills because it hides. By the time symptoms appear, the disease has often spread beyond the reach of surgery. Early detection changes everything—it transforms a death sentence into a treatable condition. The problem is that spotting the difference between a harmless spot on a chest X-ray and an actual tumor requires expertise, time, and often multiple tests. Radiologists see thousands of images. Blood work sits in a lab. Clinical notes gather in a file. These pieces of information rarely talk to each other.

Researchers in China set out to build a system that would make them speak. They assembled data from 595 patients with 618 lung lesions that had been confirmed by biopsy—the gold standard, the definitive answer. They split the cases into a training set and a test set, then built something ambitious: not one artificial intelligence model, but eleven of them, each using a different mathematical approach. Some models specialized in pattern recognition. Others excelled at weighing competing signals. The researchers then stacked these eleven models on top of each other, creating a voting system where the strongest performers had the most influence.

The key innovation was refusing to let the models work in isolation. Instead of asking a radiologist to read an image, or a lab technician to interpret blood markers alone, the system fused everything together. It looked at the shape and size of the lesion on imaging. It measured how much the tumor touched the pleura, the membrane surrounding the lungs. It factored in the patient's age. And it incorporated blood markers—proteins that tumors release into the bloodstream, like CEA and CYFRA21-1, which can signal malignancy.

When tested on cases the system had never seen before, the best-performing model—a combination called Stacking-XGBoost—achieved an accuracy score of 0.83 on a scale where 1.0 is perfect. More importantly, it caught more cancers. When the researchers compared it to a simpler model using only imaging, the combined approach reduced missed diagnoses. In the test set, false negatives dropped from 17 cases to 13. That difference matters. A missed diagnosis means a patient walks out of the clinic thinking they're fine when they're not.

But accuracy alone doesn't win trust in a hospital. Doctors need to understand why the machine made its recommendation. If a system says a lesion is malignant, the clinician needs to know: Was it the size? The location? The blood marker? Using a technique called SHAP—a method for interpreting machine learning decisions—the researchers opened the black box. They found that a standardized imaging scoring system called LUNG-RADS was the single most important factor across all models. Imaging features like pleural contact area consistently pushed the system toward a cancer diagnosis. Blood markers like CYFRA21-1 reinforced that signal.

The external validation tested the system on 126 lesions from 118 new patients it had never encountered. Performance varied across the different models, but the ensemble approaches held up reasonably well. The ExtraTrees base model achieved an accuracy of 0.78. The Stacking-GBM ensemble reached 0.76. These numbers matter because they suggest the system could generalize beyond the original training data—a critical requirement for any tool meant to work in real hospitals with real patients.

What emerges from this work is a template for clinical AI that doesn't ask doctors to blindly trust a number. By combining multiple data streams and then explaining which signals mattered most, the researchers created something that could actually fit into a radiologist's workflow. The system doesn't replace human judgment. It augments it, flagging high-risk cases and showing its reasoning. For a disease where early detection can mean the difference between five years and fifty, that kind of partnership between human expertise and machine precision might save lives.