AI Reads Plain CT Scans to Tell Benign Lung Nodule from Early Cancer

When a routine chest scan reveals a solid nodule deep in the lung, patients and their doctors face one of the most consequential questions in modern medicine: is this a harmless quirk of biology or the first sign of an invasive cancer? A research team in Kunming, China, has now tackled a particularly stubborn version of that dilemma, using artificial intelligence to squeeze diagnostic information out of ordinary, contrast-free CT images. Their findings, published in BMC Medical Imaging, suggest that machine-learning analysis of unenhanced CT scans can help separate a benign tumor called pulmonary hamartoma from a dangerous lookalike, invasive lung adenocarcinoma, though the results come with important statistical caveats that the authors themselves are careful to flag.

The diagnostic challenge at the heart of the study is a genuine clinical puzzle. Pulmonary hamartomas are benign growths made up of disorganized mixtures of cartilage, fat, and connective tissue, and they are among the most common benign tumors found in the lung. The classic teaching is that hamartomas announce themselves with telltale features: popcorn-like calcifications or visible fat deposits on a CT scan. But not every hamartoma follows the script. A subset, known as enhancing noncalcified pulmonary hamartomas, or eNCPHs, show neither calcification nor fat, and instead light up with contrast enhancement just like a malignant tumor would. When such a lesion sits in the outer, peripheral lung and appears as a solid mass, it can be virtually indistinguishable on routine imaging from peripheral solid invasive lung adenocarcinoma, the most common form of lung cancer. The only definitive answer has traditionally been a biopsy or surgical resection, procedures that carry cost, risk, and anxiety.

Radiomics offers a way around this impasse. The technique involves extracting hundreds of quantitative features from medical images, features that capture texture, shape, and intensity patterns invisible to the naked eye. A hamartoma’s cartilage content and a cancer’s chaotic cellular architecture leave different statistical fingerprints in the pixel data, even when the two lesions look identical to a radiologist. Crucially, the Chinese team focused on unenhanced CT scans, the images taken before contrast dye is injected. If reliable features could be extracted from these plain scans, the approach could potentially work even for patients who cannot receive contrast agents, such as those with kidney impairment or contrast allergies, and it would sidestep the need for an additional enhanced scan.

The study design reflects a growing awareness in the machine-learning community that how you validate a model matters as much as how you build it. The researchers assembled 95 pathologically confirmed nodules, 42 enhancing noncalcified pulmonary hamartomas and 53 invasive adenocarcinomas, all identified from patients treated between May 2020 and January 2026. Every patient gave telephone informed consent, and the protocol was approved by the Medical Ethics Committee of the First Affiliated Hospital of Kunming Medical University. The cases were then split at the patient level using a stratified random 7:3 ratio, with 66 nodules assigned to the model development set and 29 reserved for a second-look internal evaluation. Importantly, none of the final model’s parameters were tuned on those 29 cases, which had been examined in an earlier analysis but never used to optimize the frozen model.

Before any machine learning began, the team imposed a reproducibility gate. Radiomics features are notoriously sensitive to how images are acquired and processed, so the researchers filtered out features that failed to show acceptable agreement, measured using intraclass correlation coefficients. They then reduced the feature set using a training-only pipeline, meaning the feature reduction was performed exclusively on the development data to avoid information leaking into the test set. Five different classifier families were compared, including logistic regression, support vector machines, Gaussian naive Bayes, linear discriminant analysis, and stochastic gradient descent methods. The winner of the model-selection contest was logistic regression, built on just eight features, and its performance was assessed through a rigorous repeated 5-fold by 10 nested cross-validation scheme, a computationally demanding procedure that provides a more honest estimate of how a model will behave on unseen data than a single train-test split.

The headline numbers tell a story of moderate but real diagnostic power. In the full-development model-selection cross-validation, logistic regression achieved a mean area under the receiver operating characteristic curve, or AUC, of 0.786, though with considerable variability across repeated splits. When the frozen models were applied to the 29 reserved cases, the results were more granular. A model based on traditional morphologic features alone achieved an AUC of 0.697. Standalone radiomics reached 0.769. But the most interesting result came from a stacked model, which combined the radiomics probabilities with a separate piece of clinical information: the presence or absence of pleural indentation, a radiologic sign where a tumor pulls on the pleural membrane lining the lung. That combined model achieved an AUC of 0.827, with a 95 percent confidence interval of 0.665 to 0.989. In practical terms, the stacked model reached 87.5 percent sensitivity, 69.2 percent specificity, and 79.3 percent accuracy on the internal test set.

Yet the authors are refreshingly candid about what these numbers do not prove. Pairwise DeLong comparisons, a statistical test for whether one AUC is significantly better than another, showed no significant differences between the three models after Holm adjustment for multiple comparisons, with all adjusted P values at or above 0.150. In other words, while the stacked model had a numerically higher AUC, the study could not establish that it offered genuine incremental value over the simpler alternatives. The small size of the second-look evaluation, just 29 cases, means the confidence intervals are wide, and the true performance of any of these models could be substantially better or worse than the point estimates suggest. Feature selection itself proved unstable across repeated nested splits, meaning different runs of the analysis could pick different sets of features, a warning sign for anyone hoping to identify a universal radiomic signature.

Perhaps the most sobering finding concerns calibration, a concept that often gets less attention than discrimination but matters enormously for clinical use. Discrimination, measured by AUC, tells you whether a model ranks cancer cases higher than benign cases. Calibration tells you whether the model’s actual probability outputs can be trusted as absolute diagnostic probabilities. The standalone radiomics model had a calibration slope of just 0.026, with a confidence interval spanning from negative 0.105 to 0.157. A perfectly calibrated model would have a slope of 1.0. A slope near zero means the model’s probabilities are severely miscalibrated, essentially bearing no relationship to true diagnostic likelihood. The authors explicitly warn that the standalone radiomics probabilities should not be interpreted as absolute diagnostic probabilities, a caveat that any clinician hoping to use such a tool in practice must take seriously.

The study also confronts the reality of acquisition heterogeneity, the fact that CT scans in a real hospital are performed on different scanners with different protocols and parameters. Radiomics features can shift when acquisition conditions change, and this variability limits how well a model trained on one set of scans will transfer to another. The authors note that this heterogeneity, alongside the small evaluation set, the severe miscalibration, and the feature-selection instability, means external validation on independent datasets from other institutions is required before any clinical use can be contemplated. They have taken an unusually transparent step to support that future work: they released a frozen model reproducibility package containing the author-verified PyRadiomics parameter file, all 50 fold-specific stacked-model pipelines, human-readable parameters, executable prediction code, software requirements, checksums, and a verification report, allowing other researchers to reproduce their exact results.

What emerges from this study is a portrait of medical AI research at its most honest and most careful. The technology shows genuine promise: unenhanced CT radiomics can achieve moderate discrimination between a benign hamartoma and an invasive adenocarcinoma within a carefully defined contrast-enhancing phenotype, potentially sparing some patients unnecessary invasive procedures. The stacked model incorporating pleural indentation hints that combining imaging-derived features with clinical signs could push performance higher. But the researchers resist the temptation to oversell. They acknowledge that their sample was small, that their validation was internal rather than external, that their probabilities were miscalibrated, and that their feature selection was unstable. In a field often criticized for hype, that restraint is itself a kind of scientific achievement, and it sets a clear roadmap for the external validation studies that must come before these tools reach the clinic.

Subject of Research: Unenhanced CT radiomics within a contrast-defined enhancing phenotype for differentiating noncalcified pulmonary hamartoma from peripheral solid invasive lung adenocarcinoma: a retrospective development and internal test study

Article Title: Unenhanced CT radiomics within a contrast-defined enhancing phenotype for differentiating noncalcified pulmonary hamartoma from peripheral solid invasive lung adenocarcinoma: a retrospective development and internal test study

Article References: Zhang, X.-Y., Tang, C.-H., Kang, X.-M., Yan, J., Xu, Y.-M., Lu, L., Long, Y., & Duan, H. (2026). Unenhanced CT radiomics within a contrast-defined enhancing phenotype for differentiating noncalcified pulmonary hamartoma from peripheral solid invasive lung adenocarcinoma: a retrospective development and internal test study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02933-9

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02933-9

Keywords: Unenhanced, radiomics, within, contrast-defined, enhancing, phenotype, differentiating, noncalcified, pulmonary, hamartoma, peripheral, solid