The study, published in BMC Medical Imaging, was led by Guohua Wang and Huiyu Kuang, who contributed equally as first authors, together with colleagues at the Guangdong Provincial Hospital of Integrated Traditional Chinese and Western Medicine and the Neusoft Institute Guangdong. It was a retrospective, single-center investigation of 253 patients whose ground-glass nodules had been surgically removed and examined under a microscope. The pathological diagnoses spanned the entire biological continuum of lung adenocarcinoma: 19 lesions were atypical adenomatous hyperplasia, a pre-cancerous change; 34 were adenocarcinoma in situ, a cancer confined to its original location; 47 were minimally invasive adenocarcinoma; and 153 were fully invasive adenocarcinoma, the stage at which tumor cells have breached surrounding structures and gained the ability to spread.

What makes this spectrum so difficult for conventional imaging is that the lesions all look superficially similar on a CT scan. A ground-glass nodule is, by definition, a hazy area that does not obscure the underlying blood vessels and airways. Whether that haze represents a lazy cluster of atypical cells or an aggressive invasive tumor can depend on subtleties far below the threshold of human visual perception, such as the statistical texture of pixel intensities, the fine structure of the interface between the nodule and normal lung, and three-dimensional shape characteristics that no radiologist can reliably quantify by eye. Radiomics was designed precisely to close that gap. The approach converts a medical image into thousands of quantitative features, each capturing some measurable property of the tissue, and then uses statistical and machine-learning methods to find which of those properties actually carry diagnostic information.

In this study, the workflow began with manual three-dimensional delineation of each nodule on thin-section CT scans, meaning that radiologists traced the lesion slice by slice to build a complete volumetric region of interest. From each of these volumes the team extracted 1,836 candidate radiomic features. That enormous feature list was then subjected to a disciplined sequence of dimensionality reduction. Features with zero variance across the cohort were discarded first, followed by univariate screening to remove features with little relationship to the outcome, Pearson correlation filtering to eliminate redundant features that duplicated one another’s information, and finally least absolute shrinkage and selection operator regression, a technique known as LASSO that shrinks irrelevant coefficients to zero and keeps only the most predictive variables. The surviving features, just 10 for the first classification task, 14 for the second, and 10 for the third, were fed into a random forest classifier, an ensemble method that combines many decision trees to produce a robust final prediction.

The researchers structured the problem as three binary tasks, each defined by a pathological threshold along the adenocarcinoma continuum. Task 1 asked whether a nodule was atypical adenomatous hyperplasia or something further along the spectrum. Task 2 separated the two earliest, non-invasive categories from the two invasive ones. Task 3, arguably the most clinically consequential, distinguished invasive adenocarcinoma from everything less threatening. The 253 patients were randomly divided in a 4:1 ratio, stratified by pathological class, into a training set of 202 cases and a held-out internal test set of 51 cases, ensuring that the models were evaluated on lesions they had never seen during learning.

The results were impressive, particularly for the task that matters most in the clinic. On the held-out test set, the radiomic model achieved an area under the receiver operating characteristic curve of 0.9016 for distinguishing atypical adenomatous hyperplasia from all other lesions, 0.8898 for separating non-invasive from invasive disease, and 0.9484 for identifying invasive adenocarcinoma. An AUC of 0.95 means that in roughly 95 percent of cases, a randomly chosen invasive cancer would be scored as more suspicious than a randomly chosen non-invasive lesion. For that third task, the model reached an accuracy of 0.8824, with sensitivity of 0.8387, meaning it caught the great majority of true invasive cancers, and specificity of 0.9500, meaning it rarely raised a false alarm about lesions that were actually indolent.

As a secondary, exploratory analysis, the team also tested whether features borrowed from deep learning could match the handcrafted radiomic approach. They used a ResNet18 network pretrained on the ImageNet database, a standard transfer-learning strategy in which a neural network trained on millions of everyday images is repurposed for medical tasks. The network generated 512-dimensional feature vectors for each nodule, which were then selected by LASSO and classified by random forest in the same manner as the radiomic features. The pooled five-fold cross-validation AUCs for this deep-learning pipeline were 0.8520, 0.8203, and 0.8338 across the three tasks, respectable but consistently below the primary radiomics results. The authors were careful to note that this comparison was not entirely fair: the deep-learning analysis used display-export images rather than original imaging data, applied a non-nested feature selection procedure, and followed a validation framework different from the primary analysis, so it should be regarded as hypothesis-generating rather than definitive.

The clinical implications of the primary findings are considerable. Today, the management of ground-glass nodules is a delicate balancing act. Surveillance with repeated CT scans is recommended for small, indolent-appearing lesions, because many of them never progress and surgery carries real risks. But the longer a truly invasive cancer is left in place, the greater the chance of progression. A validated radiomic tool could, in principle, give clinicians an objective, quantitative estimate of invasiveness at the moment of the first scan, helping to decide who can safely be watched and who should be referred for surgical resection. It could also spare patients with benign lesions the psychological burden and cumulative radiation exposure of years of follow-up imaging, and reduce the number of operations performed on nodules that would never have harmed them.

Yet the authors themselves are emphatic that these results do not yet establish clinical utility, and their caveats deserve as much attention as the headline numbers. The study was conducted at a single center, raising questions about whether the models would generalize to different scanners, reconstruction protocols, and patient populations. The atypical adenomatous hyperplasia subgroup contained only 19 patients, just 4 in the test set, which makes the Task 1 confidence intervals wide and the estimates correspondingly fragile. The class distribution was heavily imbalanced, with invasive adenocarcinoma accounting for more than 60 percent of cases, a situation that can inflate apparent performance. And because every lesion in the cohort was surgically resected, the study population is inherently skewed toward nodules that were already considered suspicious enough to operate on, leaving out the many ground-glass nodules managed conservatively in the real world. External validation in independent, multi-center cohorts will be essential before any of these models can approach the clinic.

Even with those limitations, the study adds to a rapidly growing body of evidence that the information needed to characterize a lung lesion is already present in images clinicians take every day, waiting to be decoded. Radiomics sits at a pragmatic middle ground between purely visual interpretation and heavyweight deep learning: its features are mathematically defined, traceable to specific image properties, and explainable through tools such as SHAP analysis, which the authors employed to understand which features drove their models’ decisions. As validation efforts expand and the field moves toward standardized feature definitions and multi-center trials, the prospect of a routine CT scan doubling as a non-invasive biopsy, one that tells a patient with a hazy lung nodule whether they need a surgeon or simply a follow-up appointment, moves steadily closer to reality. For the millions of people each year who learn they have a ground-glass nodule, that future cannot come soon enough.

Subject of Research: CT radiomics for preoperative stratification of pulmonary adenocarcinoma-spectrum lesions presenting as ground-glass nodules

Article Title: CT radiomics for stratifying pulmonary adenocarcinoma-spectrum lesions presenting as ground-glass nodules: a single-center retrospective study

Article References: Wang, G., Kuang, H., Wu, W., Wu, Y., Gao, L., Feng, W., Tang, D., Lin, Y., Huang, X., & Zhu, L. (2026). CT radiomics for stratifying pulmonary adenocarcinoma-spectrum lesions presenting as ground-glass nodules: a single-center retrospective study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02907-x

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02907-x

Keywords: CT radiomics, ground-glass nodule, lung adenocarcinoma, invasive adenocarcinoma, machine learning, random forest, LASSO, transfer learning, ResNet18, medical imaging, lung cancer screening, BMC Medical Imaging