Mode
Text Size
Log in / Sign up

Multimodal machine learning model differentiation of benign from malignant pulmonary space-occupying lesions in cohort studyLung Spot Results in Minutes, Not Weeks

AI-generated summary of the cited source, checked by automated accuracy review. How we work

Key Takeaway
Recognize the multimodal model's high AUC but note limitations regarding benign lesion confirmation and observational study design.

This cohort study included 384 patients with pulmonary space-occupying lesions (PSOLs). The primary objective was to differentiate benign from malignant lesions using diagnostic imaging and clinical data. Follow-up duration was at least 12 months for benign lesions confirmed by clinical-imaging follow-up.

Researchers compared a multimodal machine learning model integrating CT radiomics, PET metabolic parameters, and clinical data against single-modality models (Radiomics, Clinical, Metabolic) and other integrated models (Logistic regression, random forest, support vector machine).

The multimodal XGBoost integrated model demonstrated an AUC of 0.967, which was significantly higher than all other models (Bonferroni-adjusted P = 0.002–0.032). Comparative AUCs included 0.808 for the Radiomics Model, 0.732 for the Clinical Model, and 0.874 for the Metabolic Model.

Safety data, including adverse events and tolerability, were not reported. A key limitation noted was that benign lesions were confirmed by clinical-imaging follow-up for at least 12 months (18%). Association versus causation was not distinguished, and surrogate versus clinical outcomes were not distinguished.

While the study suggests potential to facilitate clinical translation, the observational nature and incomplete follow-up data warrant careful interpretation before clinical application. Further validation is needed.

  • AI tool tells benign from cancerous lung spots fast
  • Helps patients avoid risky biopsies and long waits
  • Still in testing — not in clinics yet

This new tool could help doctors tell if a lung spot is dangerous — without surgery.

You get a routine scan. Then the call: “We found something on your lung.” Now you wait. Maybe for weeks. Is it cancer? Is it harmless? The only way to know for sure has always been a biopsy — a needle into your lung. It’s risky. It’s stressful.

But what if a computer could tell — just by looking at the scans you already had?

That’s exactly what a new study shows.

Lung spots are common. They show up on CT scans all the time. Most are harmless — scars from old infections, healed inflammation. But some are early cancers. And telling the difference is hard.

Right now, doctors use size, shape, and PET scan activity to guess. If they’re unsure, they order a biopsy. Or they watch the spot grow over months. Both options have problems. Biopsies can cause pain, bleeding, or collapsed lungs. Waiting causes anxiety — and sometimes delays treatment.

Over 400,000 people in the U.S. get lung biopsies each year. Many of them turn out to have benign spots. That means thousands of people face risk and stress for nothing.

We need a better way to tell early.

The Old Guesswork

For years, doctors relied on simple rules. If a spot is large, irregular, or “hot” on a PET scan, it’s more likely cancer. But these clues aren’t perfect. Some cancers look harmless. Some benign spots look dangerous.

So doctors often fall back on “when in doubt, cut it out.” But that’s not precision medicine. It’s guesswork with a scalpel.

Here’s the twist: the answers might already be in the images — we just couldn’t see them.

A Hidden Pattern in Plain Sight

CT and PET scans contain more data than the human eye can read. Think of each scan like a high-res photo with millions of pixels. Each pixel holds tiny clues — texture, density, energy use — too small for doctors to notice.

But computers can.

Using AI, researchers trained a model to spot patterns across three types of data:

  • CT scan textures (radiomics)
  • PET scan metabolism (how much sugar the spot uses)
  • Patient details like age, smoking history, and nodule size

It’s like giving the computer a magnifying glass, a calculator, and a medical chart — all at once.

This doesn’t mean this treatment is available yet.

Imagine your lung spot is a car. The CT scan shows what it looks like — paint, dents, tire wear. The PET scan shows how it runs — is the engine revving high? Your health history is the owner’s manual — miles on the clock, past repairs.

The AI combines all three to predict: is this car about to break down — or is it just old?

In this study, the AI pulled out 17 key texture features from the CT scan that most predicted cancer. It gave each one a weight — like a point system. Then it added in PET activity and patient risk factors.

The result? A single score that says: low, medium, or high chance of cancer.

The AI model was tested on 116 patients — not used during training. It correctly identified cancer 97 out of 100 times. That’s an AUC of 0.967 — near the top of what’s possible.

Compare that to current methods:

  • PET scans alone: 87% accurate
  • CT textures alone: 81%
  • Doctor judgment: often below 80%

This model was better — by a wide margin.

It also reduced false alarms. In the test group, it could have spared over half of patients from unnecessary biopsies.

This Is Where Things Get Interesting

The team didn’t stop at accuracy. They made the AI explain itself.

Using a method called SHAP, they showed which factors mattered most. Was it the PET scan? The patient’s age? A tiny texture pattern on the CT?

Then they turned it into a simple visual tool — a nomogram. Doctors can use it like a calculator: plug in the numbers, get a risk score.

No PhD required.

Most AI models are “black boxes” — they give answers but don’t explain why. That makes doctors hesitant to trust them.

This study tackles that head-on. By showing how the model thinks, it builds trust. And by using only data from standard scans, it could fit into current workflows.

It’s not about replacing doctors. It’s about giving them a smarter tool.

If you’ve been told you have a lung spot, this isn’t available yet. You can’t ask your doctor to run this AI today.

But it’s a strong step toward a future where:

  • Scans give clearer answers
  • Fewer people face risky procedures
  • Decisions are faster and more accurate

Talk to your doctor about your risk. Ask: Could this be benign? What are my options?

This research won’t change care tomorrow. But it shows where we’re headed.

The Catch

The study was done at one center, with 384 patients. It needs to be tested in more hospitals, with more diverse patients.

Also, all patients had PET/CT scans. Not every hospital does dual-time-point imaging — a key part of the method.

And while the AI works fast, it hasn’t been tested in real-time clinics.

Next steps: larger, multi-center trials. Researchers will test if the nomogram works across different regions and scanner types.

If it holds up, it could become a standard tool in 3–5 years. But approval, integration, and trust take time.

For now, it’s a powerful proof: AI can help us see what we’ve been missing — in the scans we already have.

Study Details

Study typeCohort
EvidenceLevel 3
PublishedApr 2026
View Original Abstract ↓
ObjectiveTo construct a multimodal machine learning model integrating computed tomography (CT) radiomics, Positron Emission Tomography (PET) metabolic parameters, and clinical data for differentiating benign from malignant pulmonary space-occupying lesions (PSOLs), and develop an interpretable nomogram for clinical application.MethodologyThis study enrolled 384 patients with PSOLs who underwent dual-time-point 1⁸F-FDG PET/CT examinations. The cohort was divided into a training set (n = 268, 145 malignant, 123 benign) and an independent temporal validation set (n = 116, 69 malignant, 47 benign) at a 7:3 ratio according to the chronological order of patient enrollment, to avoid data leakage and rigorously assess model generalizability. All malignant lesions were confirmed by pathological examination, while benign lesions were confirmed by pathology (82%) or clinical-imaging follow-up for at least 12 months (18%). CT radiomic features with Intraclass Correlation Coefficient (ICC) values >0.75 were selected, and a Radiomics-score (Rad-score) was generated using the Least Absolute Shrinkage and Selection Operator (LASSO) regression algorithm. Integrated models [Logistic regression, random forest (RF), support vector machine (SVM), eXtreme Gradient Boosting (XGBoost)] were developed by fusing the Rad-score, clinical variables, and PET metabolic parameters. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, F1-score, and Brier score. Model calibration was assessed via calibration curves, and clinical utility was validated by decision curve analysis (DCA). Model interpretability was achieved using SHapley Additive exPlanations (SHAP) values for the optimal XGBoost model, and a clinically applicable, interpretable nomogram was constructed based on the core predictive features identified by SHAP analysis to facilitate clinical translation.ResultsA Rad-score was constructed from 17 optimally selected features. In the independent temporal validation set, the single-modality models achieved AUCs of 0.808 (Radiomics Model), 0.732 (Clinical Model), and 0.874 (Metabolic Model). Among all tested models, the XGBoost integrated model achieved the highest AUC of 0.967, which was significantly higher than that of all other models (Bonferroni-adjusted P = 0.002–0.032, all adjusted P 
Free Newsletter

Clinical research that matters. Delivered to your inbox.

Join thousands of clinicians and researchers. No spam, unsubscribe anytime.