Preoperative prediction of pulmonary ground-glass nodule infiltration status by CT-based radiomics combined with neural networks

Abstract

Objective

The infiltration status of pulmonary ground-glass nodules (GGNs) exhibits significant variability, demanding tailored surgical strategies and individualized postoperative adjuvant therapies. This study explored the preoperative assessment of GGN infiltration status using computed tomography (CT) imaging integrated with a neural network to enhance the precision of clinical decision-making in surgical planning and therapeutic interventions.

Methods

This multicenter retrospective study analyzed clinical data to quantify mismatch rates in surgical approaches across varying infiltration statuses. Regions of interest (ROIs) within the CT lung window level were manually delineated using ITK-SNAP software, enabling the extraction of relevant CT imaging features, including morphological descriptors, first-order statistical parameters, texture attributes, and high-order characteristics. Feature selection was performed using the Lasso algorithm to identify the most predictive variables, which were subsequently incorporated into the radiomics-based neural network model. The neural network architecture combined a 3D convolutional neural network (CNN) with random rotations for data augmentation and employed pre-trained parameters to optimize model weights.

Results

The radiomics-integrated neural network exhibited high predictive performance, achieving an area under the subject operating characteristic curve (AUC) of 0.85, with validation set AUCs of 0.66 and 0.71. Additionally, the predicted mismatch rate between lobectomy and sublobectomy was 21.48%, representing a 35.57% reduction, while the mismatch rate within sublobectomy decreased by 13.66%, reaching 10.73%

Conclusion

The neural network-enhanced imaging model provides a robust predictive tool for assessing the preoperative infiltration status of pulmonary GGNs. Its application significantly reduces mismatch rates in surgical decision-making, contributing to more precise and individualized treatment strategies.

Clinical perspectives

The application of radiomics in cancer research has been extensively explored, yet its integration into clinical surgical guidance remains limited. This study employs deep learning combined with radiomics to predict patients' infiltration statuses, optimizing preoperative surgical planning with greater precision. By refining surgical strategies, this method supports more tailored treatment approaches, ultimately improving patient outcomes.

Introduction

Lung cancer remains a leading cause of cancer-related mortality worldwide, with both incidence and mortality rates increasing in China [1]. Major risk factors include smoking and air pollution, while gender and family history contribute to susceptibility [2,3,4]. The adoption of low-dose computed tomography (LDCT) for lung cancer screening has expanded, particularly in response to the COVID-19 pandemic, and it is now a recommended screening modality in China [5]. Despite its widespread use, LDCT presents challenges, including difficulties in differentiating indeterminate lung nodules and the risk of overdiagnosing indolent tumors [6,7,8,9]. Radiomics, an advanced imaging analysis technique that extracts quantitative features from modalities such as CT, has demonstrated potential for enhancing diagnostic precision in lung cancer assessment [10, 11].

Ground-glass nodules (GGNs) are categorized based on their compositional elements and radiological features into three distinct types: (1) Pure ground-glass nodules (pGGNs), exhibiting homogeneous ground-glass opacity without solid components; (2) Mixed ground-glass nodules (mGGNs), characterized by a combination of ground-glass regions and solid elements; and (3) Solid nodules, which, despite falling outside the GGN classification, warrant consideration due to the clinical relevance of their solid components in lung nodule assessment. The extent and distribution of solid components are key determinants in evaluating lung nodules, particularly in mGGNs. Notably, solid nodules represent a separate classification of pulmonary nodules, distinct from GGNs, which are defined by their ground-glass density—a feature absent in solid nodules [5]. Current guidelines recommend surveillance intervals based on nodule size and classification [5]. However, existing approaches for assessing GGN infiltration status remain limited, particularly in preoperative settings. Traditional radiomics methods rely on handcrafted features, whereas deep learning models often demand extensive datasets and suffer from limited interpretability [12, 13]. Additionally, limited research has explored the integration of radiomics and neural networks for preoperative GGN infiltration prediction.

To address these challenges, this study presents a novel framework integrating CT-based radiomics with advanced neural networks to predict the infiltration status of GGNs. By utilizing the complementary strengths of these methodologies and incorporating clinical parameters, the model achieves improved predictive precision and robustness. A systematic comparison with various machine learning and deep learning models further delineates the superiority of the proposed approach in predictive performance and generalizability. Our model not only improves the safety and effectiveness of surgery, but also provides more precise and individualized treatment options for patients with less aggressive lung adenocarcinoma.

Materials and methods

Patients

This study was conducted as a multicenter retrospective analysis. The study's workflow was illustrated in Fig. 1. Clinical data were obtained from patients undergoing GGN surgeries at Changzhou First People’s Hospital between June 2019 and June 2020, with 354 cases designated as the model test set. Additionally, data from 54 patients treated at Zhongshan Hospital Affiliated with Fudan University and 32 patients from Changzhou Second People’s Hospital were compiled as a validation set. The inclusion criteria consisted of the following: (1) Patients who underwent lung cancer surgery; (2) Pathological confirmation of atypical adenomatous hyperplasia (AAH), adenocarcinoma in situ (AIS), minimally invasive adenocarcinoma (MIA), or invasive adenocarcinoma (IA); (3) Availability of CT imaging data with a slice thickness below 2 mm from Changzhou First People’s Hospital; (4) GGN diameter under 2 cm; and (5) A solid component proportion in GGNs not exceeding 50%, with only solitary ground-glass pulmonary nodules meeting this criterion included. Exclusion criteria included: (1) Deceased patients; (2) Individuals with severe cardiovascular or cerebrovascular diseases or abnormal liver and kidney function; (3) Patients diagnosed with COPD; (4) Cases of large cell lung cancer or small cell lung cancer; (5) Central lung cancer; and (6) Mixed lung cancer subtypes.

Fig. 1

Research process

CT scanning parameters

CT scans were performed as a standard component of the routine testing protocol. Thin-slice chest CT images were obtained using equipment from multiple manufacturers, including GE, Philips, and Siemens. Notably, the external verification set also comprised images acquired from these three scanner types. The imaging parameters were set as follows: tube voltage of 140 kVp (range: 100–140 kVp), tube current of 340 mA (range: 100–752 mA), and slice thickness of 1.0 mm (range: 0.65–2.0 mm).

Tumor segmentation

After duplicating the original thin-slice CT data for all patients, the datasets were imported into ITK-SNAP, where regions of interest (ROIs) were delineated on tumor sites within the CT images. Patient information was anonymized upon data import. Two cardiothoracic surgeons, Mei Kun and Chao Ce, performed 3D ROI mapping of the entire tumor regions, with Wang Bin overseeing the process. All 3D ROI mappings were conducted using thin-slice CT images at the lung window level. To minimize inter-observer variability, both surgeons completed standardized training in tumor segmentation with ITK-SNAP before the study. A consensus meeting was subsequently convened to reconcile any discrepancies in the initial ROI delineations, ensuring consistency in the segmentation process.

Imagomics feature extraction

High-throughput features were extracted to quantitatively characterize the intrinsic properties of the ROI. Lung images were processed using an imaging omics feature extractor (Refer to http://epub.cnipa.gov.cn/Dxb/IndexQuery, patent number: CN116542937A), including multiple feature categories, including shape descriptors, first-order statistical metrics, texture patterns, higher-order computational attributes, and model-derived transformations. Each category was independently extracted from CT images, establishing a systematic basis for quantitative tumor analysis.

Deep learning feature extraction

CT image features were extracted using 3D convolutional neural networks, with CT scans augmented through random rotations at varying angles. To ensure consistency, the training and validation datasets were normalized to a range between 0 and 1. Model optimization was achieved by constructing a 3D neural network, followed by loading pre-trained weight parameters. The network architecture comprised three fully connected layers and thirteen convolutional layers. Training data were generated through multiple single-scale and multi-scale cropping operations applied to the images. Mini-batch gradient descent was then employed for iterative optimization, enabling progressive convergence toward the optimal solution. Finally, individualized CT feature outputs were generated for each patient.

Feature fusion and screening

At this stage, clinical, imaging, and neural network-derived features were integrated to construct a comprehensive feature vector for each patient.

Subsequently, the most relevant invasion-associated features were extracted from the full feature set. While tumor-related features were derived from CT images using three distinct methods, their relevance to invasion varied. To refine feature selection, the least absolute shrinkage and selection operator (LASSO) algorithm was employed during training to identify those with the highest discriminative power.

The primary objective was to distinguish between invasive and non-invasive pulmonary carcinoma through predictive modeling. The relative contribution of each feature was quantified by its assigned weight. The cost function was mathematically expressed as follows:

$$\text{J}(\text{w})=\frac{1}{\text{m}}{\sum }_{\text{i}=1}^{\text{m}}{({\text{y}}_{\text{i}}-{\text{w}}^{\text{T}}{\text{x}}_{\text{i}})}^{2}+\uplambda {\sum }_{\text{i}=1}^{\text{m}}|{\text{w}}_{\text{i}}|$$

The application of the LASSO algorithm in this study enabled the compression of numerous feature coefficients, leading to refined regression adjustments. This approach optimized feature selection by identifying variables significantly associated with infiltration. To determine the optimal regularization parameter (λ), a k-fold cross-validation strategy systematically assessed model performance across varying λ values. The parameter corresponding to the highest cross-validation performance was selected to enhance the robustness and generalizability of the feature selection process.

The traditional radiomics model extracted handcrafted features from CT images and used machine learning algorithms for classification. The CNN model utilized a deep learning architecture with multiple convolutional layers to automatically learn features from the images. The SVM model employed a kernel-based approach to classify the data based on the extracted features. By comparing these models with the proposed integrated model, we aimed to demonstrate the superiority of our approach in predicting the infiltration status of GGNs.

Classifier training

To address the imbalance in sample distribution, primarily caused by the high prevalence of infiltrative lung tumors in the collected dataset, the synthetic minority over-sampling technique (SMOTE) was applied. This method adjusts sample proportions by analyzing minority class instances and generating artificial samples to enhance dataset representativeness. By employing this approach, the training set achieved a balanced 1:1 ratio between the two data categories.

For classification and discrimination, a nonlinear support vector machine (SVM) classifier was utilized to identify tumor invasion. As a binary classification model, SVM constructs a linear decision boundary with maximal margin in the feature space. Additionally, SVM incorporates kernel functions to enable nonlinear transformations, and in this study, the radial basis function kernel was selected to enhance feature mapping.

Results

Basic clinical data

A total of 354 GGN patients were included in the validation set, comprising 149 IA, 81 MIA, 116 AIS, and 8 AAH cases. The test set was stratified into Group I and Group II based on surgical treatment approaches, with Group I including IA cases and Group II comprising MIA, AIS, and AAH cases. According to the lung cancer diagnosis and treatment guidelines issued by the Chinese Medical Association, lobectomy is the standard surgical procedure for Group I, whereas sublobectomy is recommended for Group II. However, both groups in this study exhibited discrepancies between the recommended and actual surgical interventions, quantified as the Mismatch Rate (Fig. 2). Group I demonstrated a mismatch rate of 57.05%, while Group II reported 24.39%. Comprehensive data were summarized in Tables 1, 2, 3 and 4, with no statistically significant differences observed between the variables in the training and validation sets (P > 0.05).

Fig. 2

Mismatch programming

Table 1 Patient data of Changzhou First People’s Hospital

Table 2 Patient data of Zhongshan Hospital Affiliated to Fudan University

Table 3 Patient data of Changzhou Second People’s Hospital

Table 4 Surgical mismatch rates

Tumor marker

Thin-slice CT scans were imported into ITK-SNAP for tumor localization (Fig. 3), with labeling performed independently by two cardiothoracic surgeons. Discrepancies in annotations were adjudicated by a senior cardiothoracic surgeon with 15 years of clinical experience. The interobserver agreement, assessed through interclass correlation coefficients, ranged from 0.92 to 0.97, demonstrating strong concordance.

Fig. 3

Tumor ROI profiling

Feature extraction and model construction

Tumor site features labeled in CT data were extracted, with a total of 624 features screened using radiomics and deep learning algorithms. Among them, 112 had specific interpretability, while 512 lacked clear clinical significance (Supplementary Table 1). Despite the comprehensive extraction of tumor characteristics from CT images via both methodologies, no substantial association with invasion was identified. To refine feature selection, the LASSO algorithm was applied to pinpoint those with the highest classification performance within the training set. Following selection, optimal regularization parameters, essential to the LASSO model, were determined within the range [10⁻³, 10¹], and corresponding feature weights were generated. Features assigned higher weights were selected as classifier inputs. The optimization process for regularization parameters was illustrated in Fig. 4, while the detailed LASSO algorithm workflow was presented in Fig. 5. Three predictive models were developed using radiomics, deep learning convolutional networks, and their integration. Comparative analysis demonstrated that the combined approach yielded the highest accuracy (Table 5).

Fig. 4

Lasso algorithm filtering

Fig. 5

Feature selection

Table 5 Comparison of three models

Model application and effectiveness evaluation

Comparative analysis of the three models revealed that the radiomics approach integrated with the neural network exhibited superior predictive performance in evaluating tumor myographic infiltration (Fig. 6). The model’s accuracy was further validated using datasets from three hospitals (Table 6), yielding an AUC of 0.85 in the training cohort, 0.71 in validation cohort 1 (Changzhou Second People’s Hospital), and 0.66 in validation cohort 2 (Zhongshan Hospital Affiliated to Fudan University). Concurrently, the model estimated the clinical mismatch rate during lung cancer surgery. According to its projections, Group I demonstrated a mismatch rate of 21.48%, representing a 35.57% reduction, while Group II recorded a mismatch rate of 10.73%, indicating a 13.66% decrease.

Fig. 6

Diagram of a 3D neural network model. A Model flow chart. B Model internal structure diagram

Table 6 Verification of model performance

Model application and effectiveness evaluation

To further assess the proposed model’s efficacy, its performance was evaluated against several widely used machine learning and deep learning models in medical imaging, including traditional radiomics models, convolutional neural networks, and SVMs. Comparative analysis demonstrated that the integrated model outperformed conventional approaches in predicting the infiltration status of GGNs (Table 7). ANOVA tests were conducted to compare the AUC values among four models—traditional radiomics, convolutional neural network, support vector machine, and the integrated model—to determine statistical significance. The results indicated a significant performance advantage of the integrated model over the others (P < 0.05). The integration of radiomics, neural networks, and clinical features within the proposed model yielded superior accuracy and generalizability relative to existing methodologies, highlighting its effectiveness in addressing the challenges of GGN infiltration prediction.

Table 7 Comparative evaluation of different predictive models

Discussion

As a malignancy with a high global mortality rate, tumor infiltration remains a central focus of research. Accurate preoperative assessment in early-stage patients enables the selection of individualized treatment strategies based on imaging analysis. However, conventional biopsy techniques, including CT-guided puncture and bronchoscopy, may be inadequate for non-resectable tumors. Additionally, overtreatment continues to present significant challenges in clinical practice. Statistical analysis in this study revealed a lobectomy mismatch rate of 57.05%, while sublobectomy accounted for 24.39% of unmatched cases. Such discrepancies not only expose patients to unnecessary risks but also place substantial strain on clinical workflows. To address these limitations, a predictive model integrating radiomics with neural networks was developed to enable precise and non-invasive preoperative evaluation, supporting personalized treatment planning. By incorporating radiomics and deep learning features, the proposed model enhanced the accuracy of preoperative prediction for GGN infiltration status. Improved predictive precision has the potential to significantly reduce mismatch rates in current clinical practice, optimizing surgical decision-making and improving patient outcomes. For instance, more accurate preoperative assessment could facilitate the identification of patients suited for sublobectomy, thereby preserving lung function while maintaining oncological safety.

Pulmonary GGNs can be classified into three primary pathological subtypes: AIS, MIA, and IA. AAH, which closely resembles AIS in morphology, may also be present. Prognostic outcomes vary substantially depending on the extent of infiltration. AAH can be managed through long-term observation without surgical intervention. For AIS and MIA, sublobectomy without local lymph node dissection is sufficient, yielding a 5-year survival rate of 100%. In contrast, IA requires lobectomy with local lymph node dissection, reducing the 5-year survival rate to 67% [14]. With the increasing detection of pulmonary GGNs, precise assessment of nodule infiltration is essential. However, the substantial imaging overlap among GGNs presents a significant challenge to accurate diagnosis and treatment.

In clinical practice, GGN lesion diagnosis via CT primarily relies on macroscopic imaging features visible to the naked eye, including lesion size, margin characteristics, and solid components. However, no definitive imaging marker has been established to determine the invasive state. This limitation highlights the necessity for non-invasive, cost-effective screening methods capable of efficiently identifying high-risk cases. Subtle image-based features may provide greater discriminatory power in assessing tumor invasion. In this study, 624 radiomic features were extracted and analyzed through deep learning, with the most relevant features selected via the Lasso algorithm to construct the predictive model. Results indicated that integrating radiomics with deep learning offered substantial potential for evaluating GGN infiltration status prior to biopsy or surgical intervention. The proposed model achieved an accuracy of 91%, significantly outperforming previously reported methodologies.

Accurate tumor classification in clinical practice plays a fundamental role in developing tailored and optimized treatment strategies for cancer patients. Advances in lung cancer therapy have been driven by the identification of targeted mutations and the refinement of histology-based treatment approaches [15,16,17]. For instance, pemetrexed, a chemotherapeutic agent, serves as the preferred option for stage IV lung adenocarcinoma and contributes to a higher detection rate of targeted mutations, such as EGFR mutations, which are predominantly observed in GGNs [18, 19]. The classification of tumor infiltration statuses informs the selection of appropriate surgical interventions. According to the lung cancer diagnosis and treatment guidelines issued by the Chinese Medical Association, lobectomy is recommended as the standard surgical procedure for patients with invasive state IA, whereas sublobectomy is the preferred approach for those with invasive state MIA, AIS, and AAH. Although previous studies have examined the preoperative prediction of the invasive state of lung adenocarcinoma [20], reported accuracy remains limited to 64.1%.

Cancer treatment strategies in China tend to be conservative, yet this approach often leads to overtreatment [21]. Intraoperative frozen pathological sections are routinely utilized to inform surgical decision-making for pulmonary nodules. However, in the absence of histological samples, invasive components cannot be definitively identified [21, 22]. Prior studies have reported discrepancies between intraoperative frozen section diagnoses and final histopathological outcomes [23]. These inconsistencies likely arise from the limitations in accurately assessing GGN infiltration and, to some extent, from technical constraints in performing complex segmentectomies, prompting a preference for lobectomy and contributing to overtreatment. In the present study, the discordance rate reached 57.05% for IA procedures and 24.39% for the other three GGN subtypes. The radiomics-based neural network model developed in this study demonstrated high diagnostic accuracy. Through model optimization, the discordance rate for IA procedures was reduced to 21.48%, while that for the other three GGN subtypes declined to 10.73%. These advancements enhance surgical precision and reduce overtreatment, ensuring that patients receive the most appropriate intervention based on tumor-specific characteristics.

Our model also has important clinical significance. Firstly, it minimizes the number of unnecessary lobectomies, which are more invasive and associated with higher morbidity and longer recovery times compared to sublobectomies. By accurately identifying patients who can benefit from less aggressive surgical interventions, we can preserve more lung function, enhance postoperative quality of life, and reduce the risk of long-term respiratory complications. Additionally, reducing mismatch rates can lead to substantial cost savings for the healthcare system. Lobectomies require longer hospital stays, more extensive postoperative care, and higher resource utilization compared to sublobectomies. By avoiding unnecessary lobectomies, hospitals can allocate resources more efficiently, reduce overall treatment costs, and improve the cost-effectiveness of lung cancer management. Furthermore, accurate preoperative planning can reduce the need for additional interventions or revisions, further enhancing patient outcomes and reducing the burden on healthcare resources.

The proposed integrated model was evaluated against multiple machine learning and deep learning models, demonstrating superior performance in predicting the infiltration status of GGNs. This enhanced predictive capability can be attributed to several key factors:

Integration of Radiomics and Deep Learning Features: The combination of handcrafted radiomics features and deep learning-derived representations provides a more comprehensive characterization of tumor properties. Radiomics extracts fine-grained textural and morphological details from CT images, while deep learning models capture intricate patterns that may not be readily discernible using traditional approaches. This complementary integration improves the model’s ability to differentiate between invasive and non-invasive GGNs.

Incorporation of Clinical Features: Beyond imaging-based features, the model incorporates clinically relevant data, creating a multimodal framework that strengthens predictive performance. By integrating both imaging and non-imaging variables associated with GGN behavior, the model leverages a broader spectrum of diagnostic information.

Optimized Model Architecture: The network architecture, comprising multiple convolutional and fully connected layers, is tailored to accommodate the complexity of GGN imaging data. Advanced neural network techniques, including dropout regularization and batch normalization, further enhance the model’s robustness and generalizability, mitigating overfitting and improving stability across datasets.

Integrating this model into clinical practice requires a series of essential steps. First, a multicenter prospective study should be conducted to validate its efficacy in real-world settings, necessitating data collection from a broader and more diverse patient cohort to enhance generalizability. Second, close collaboration with clinical experts and radiologists is required to refine interpretability and optimize usability. Third, continuous performance monitoring and periodic updates based on newly acquired data are imperative to maintain long-term reliability and accuracy.

In summary, this study introduces a novel framework that combines radiomics and neural networks with clinical features to predict the infiltration status of pulmonary GGNs. By leveraging multi-modal feature integration and adaptive feature selection, the approach addresses the limitations of conventional methods. The proposed model demonstrates superior predictive performance compared to other machine learning and deep learning models, as reflected by higher AUC values in both training and validation sets.

The primary contributions of this study are as follows:

• 1. **Innovative Feature Fusion**: By incorporating clinical, radiomics, and deep learning-derived features, the model effectively captures a broad spectrum of tumor characteristics, leading to improved predictive performance.

• 2. **Tailored Model Design**: The neural network design and feature selection methodology were specifically tailored to address the inherent challenges of medical imaging data, such as limited sample sizes and high-dimensional feature spaces.

• 3. **Detailed Model Comparison**: A systematic comparison with existing machine learning and deep learning models is provided, emphasizing the distinct advantages of the proposed approach.

Future work will focus on refining the model architecture and expanding the dataset to enhance robustness and generalizability. Additionally, its clinical applicability will be assessed through multicenter prospective studies to evaluate real-world effectiveness.

Despite the multicenter design of this study, the reliability of the results is constrained by the limited number of patients in the validation set, highlighting the need for a larger sample size to enhance robustness. Additionally, the clinical applicability of the model remains insufficiently validated, requiring prospective studies to assess its practical performance in real-world settings. Future research should prioritize prospective validation and optimize the model’s integration into clinical workflows to strengthen its translational potential.

Conclusion

The radiomics integrated with the neural network model developed in this study has demonstrated high efficacy in predicting the infiltration status of GGNs prior to surgery. Additionally, this model contributes to optimizing preoperative surgical planning, significantly reducing procedure mismatch rates and enabling more precise individualized treatment. Future research will focus on prospective validation studies to further assess the model’s clinical applicability and robustness in real-world settings. Moreover, expanding data integration to include genomic information will be explored to enhance predictive accuracy and provide deeper insights into the biological characteristics of pulmonary GGNs.