The incidence and mortality rates of lung cancer surpass those of all other malignant tumors, accounting for ~20% of cancer-related deaths worldwide (1). Early diagnosis and treatment are pivotal in improving the prognosis and survival rates of lung cancer patients. In the 2021 WHO classification (2), adenocarcinoma in situ (AIS) has been reclassified as a glandular precursor lesion (GPL) along with adenomatous hyperplasia (AAH), while the subcategories of lung cancer now include minimally invasive adenocarcinoma (MIA) and invasive adenocarcinoma (IAC). This classification update necessitates corresponding adjustments to clinical diagnosis and treatment protocols. Currently, management follow-up strategies predominantly apply to AAH and AIS, whereas MIA warrants prompt surgical intervention (3,4). Furthermore, although some studies have indicated comparable long-term efficacy and 5-year survival rates between AAH, AIS and MIA surgeries, others have found that MIA exhibits higher Ki67 levels and EGFR mutation rates compared with AIS (5). Therefore, achieving precise discrimination of MIA from GPL during management follow-up and pre-surgical stages provides important insights for determining optimal clinical intervention timing, implementing surgical protocols and assessing prognosis.

Previous studies have demonstrated that high-resolution low-dose computed tomography (CT) is currently the most effective screening tool, reducing lung cancer mortality by 20% (6). Over 90% of these cases represent early-stage lung cancer, with pulmonary ground glass nodules (GGNs) being the primary manifestation on CT scans. While pathological biopsy serves as the gold standard for diagnosing lung cancer types, frozen section analysis has emerged as a valuable method for rapid intraoperative assessment of nodules, distinguishing between benign and malignant lesions and determining histological subtypes. This information plays a crucial role in guiding surgical strategies for lung nodules (7,8). However, due to the small size and low density of lung nodules, sampling often yields suboptimal results, resulting in a concordance rate of only ~68% between intraoperative frozen section analysis and postoperative paraffin pathology (8). Some researchers have explored the use of traditional CT imaging features (such as nodule size, density and solid component proportion) to differentiate between MIA and GPL (9,10). However, these features exhibit significant overlap across different nodule subtypes, leading to low diagnostic efficacy (11). Moreover, the extraction of these features heavily relies on the subjective interpretation skills and clinical experience of the radiologist (8). In addition, traditional techniques often struggle to identify nodules with smaller volumes and mixed densities. The increasing prevalence of GGNs with diameters <10 mm, detected through low-dose CT screening for early lung cancer, presents new challenges in clinical diagnosis. Some studies have proposed a critical value of 10 mm diameter for distinguishing between glandular precursor lesions and invasive lesions (9,10,12). Nevertheless, clinical practice has revealed numerous GGNs ≤10mm confirmed as MIA or IAC (10), leading to continuing debates about the management strategies for GGNs of this size (13). Consequently, there is an urgent need for innovative and accurate techniques to enable the precise preoperative diagnosis of MIA and GPL.

Radiomics is a powerful technique that enables the extraction and analysis of numerous radiomics features from medical imaging data with high efficiency. This approach holds significant promise in distinguishing between different pathological subcategories of lung nodules, assessing the extent of infiltration and evaluating prognostic outcomes (14). Texture analysis, on the other hand, involves the extraction and quantitative analysis of non-macroscopic and deep-level CT image features that reflect the tumor's heterogeneity to some extent. It has demonstrated utility in various aspects, including tumor differential diagnosis, prognostic evaluation, treatment response prediction and monitoring (15–17). CT texture analysis is widely utilized for identifying benign and malignant solitary lung nodules and evaluating invasiveness, exhibiting exceptional performance (18,19). To date, there have been no reports on the application of texture analysis for the identification of MIA and GPL in sub-centimeter GGNs. Therefore, the present study aimed to develop and validate a nomogram based on CT quantitative parameters and texture features for improving the ability to discriminate MIA from AAH/AIS, thereby providing important guidance for formulating clinical treatment plans and optimizing the timing of surgical interventions.

The present study retrospectively collected surgically resected and pathologically confirmed cases of GPL and MIA that presented as sub-centimeter GGNs on CT images, then extracted and analyzed their CT quantitative and texture features. Subsequently, a nomogram model incorporating the five most informative identification indicators (SV, SVR, CT_max, CT_mean and entropy) was developed to accurately differentiate MIA from GPL. The ROC curve revealed high predictive accuracy in distinguishing between MIA and GPL, with an AUC values of 0.992 and 0.975 for the training set and validation set, respectively. Early-stage lung cancers predominantly present as solitary GGNs, posing challenges in distinguishing benign nodules from malignant ones due to the diverse manifestations and overlapping features observed on CT images. In line with previous studies by Wu et al (9,10), who extensively analyzed the CT imaging features of sub-centimeter pure GGNs, the present study identified relevant morphological features such as lesion size, vessel changes and tumor-lung interface, which reflect the invasiveness of GGNs. Moreover, previous investigations (26–28) have established a close association between the size of the solid component and the average CT attenuation of GGNs with their invasiveness and pathology. The findings of the present study corroborated these conclusions, demonstrating significant differences in SV, SVR, SQ, SQR, CT_max and CT_mean between GPL and MIA. Notably, the values of these variables were consistently higher in the MIA group compared with the GPL group. These results confirm the importance of the solid component in predicting the behavior of GGNs. To summarize, the present study retrospectively analyzed surgically resected GGNs, extracting and analyzing their CT quantitative and texture features and achieved high predictive performance in discriminating between MIA and GPL through the development of a nomogram model incorporating five informative identification indicators. The challenges associated with differentiating benign and malignant nodules in early-stage lung cancers were underscored, emphasizing the significance of considering the presence and characteristics of the solid component in predicting invasiveness and pathological behavior of GGNs.

Texture analysis plays a crucial role in efficiently and accurately extracting biological information that reflects tumor heterogeneity, which may not be discernible by visual examination of images alone. It enables quantitative assessment of subtle changes in image pixel values and their arrangement. As a result, it holds great value in qualitative diagnosis, invasiveness assessment, prognostic prediction of tumors and informed clinical management. Qiu et al (17) demonstrated the independent prognostic significance of mean CT attenuation and entropy in evaluating the invasiveness of 428 cases of clinical stage IA lung adenocarcinoma. Zhu et al (19) employed the ANOVA test and the least absolute shrinkage and selection operator algorithm to identify 18 CT texture features, including entropy. They successfully developed a diagnostic model capable of distinguishing MIA from GPL presenting as pure GGNs, achieving high identification performance with an AUC of 0.884 in the training set and 0.872 in the validation set. In the present study, there was a statistically significant difference in entropy between GPL and MIA, which is consistent with previous findings. Moreover, the combined multi-parameter model exhibited superior predictive performance, with AUC values of 0.992 in the training set and 0.975 in the validation set, surpassing the research conducted by Zhu et al (19). This disparity may be attributed to differences in the composition of enrolled cases. The present study encompassed not only pure vitreous nodules but also mixed-density nodules. Previous studies (26,28) have indicated the diagnostic relevance of the solid component within GGNs. Another study (29) employed machine learning and deep learning methods to evaluate benign and malignant pulmonary nodules, revealing AUC values of 0.763 for the support vector machine (SVM) model and 0.723 for the convolutional neural networks (CNN) model in distinguishing GPL from MIA. Notably, the predictive performance of both SVM and CNN models was lower than that of the nomogram model established in the present study. These findings suggest that CT texture analysis may offer greater assistance and efficiency in achieving this objective. The patients in the present study were sourced from a single center, resulting in a relatively small sample size, but the parameters and quality of these CT images were highly standardized and homogenized. Conversely, in a previous multicenter study (29), although the sample size was larger, challenges related to CT image homogenization may have persisted. This discrepancy could explain the superior diagnostic effectiveness of the constructed model. Consequently, future studies should focus on expanding the sample size and analyzing data from different centers to obtain a more accurate and efficient predictive model.

Nomogram models are graphical representations employed to illustrate analytical outcomes derived from multifactorial logistic regression models or Cox proportional hazards models. These models use a set of parallel, non-intersecting lines within a coordinate plane to portray the quantitative analysis diagram, depicting the functional relationship between multiple variables. By employing intuitive symbols, nomograms facilitate the calculation of disease occurrence probability, recurrence risk and prognosis. Consequently, they find extensive application in clinical research pertaining to pulmonary GGNs (30). In the present study, a nomogram model for the identification of precursor lesions and MIA was established, based on five CT quantitative and texture features, namely SV, SVR, CT_max, CT_mean and entropy. This model serves as an effective and intuitive reference standard, enabling radiologists to achieve accurate and prompt diagnoses.

In conclusion, the use of CT quantitative and texture features offered significant utility in the differentiation of MIA from GPL. The nomogram model developed demonstrated superior discriminatory capabilities. This model's diagnostic efficacy, combined with its graphical representation, facilitates the precise classification of GGN types by radiologists, thereby aiding clinicians in making informed decisions regarding treatment and follow-up strategies for GGNs. Nonetheless, the present study has certain limitations that require acknowledgment. It was a retrospective, single-center study with a small sample size, which may introduce selection bias and potentially affect the precision and applicability of the model. In the future, efforts should be made to enhance the model's robustness and applicability by increasing the sample size, conducting multi-center studies and incorporating external test sets.