Radiomics features combined with clinical and CT features for predicting the Ki-67 index in stage T1 non-small cell lung cancer patients: a multicenter study

Introduction

The high death rate and difficulty in detecting the disease until it is at its late stages make non-small cell lung cancer (NSCLC) a major public health concern worldwide (1). The Ki-67 index, which measures cellular proliferation, is an important tool for gauging tumour aggressiveness and the effectiveness of treatments (2). While the Ki-67 index is not part of the formal diagnostic criteria, it plays a crucial role in personalized medicine, enabling tailored therapeutic strategies on the basis of specific tumour characteristics (3,4). At present, surgical histology and biopsies are the only methods for determining Ki-67 expression in lung cancer. However, these methods have their limitations, such as the use of tiny samples not necessarily being representative of the tumour’s heterogeneity (5,6). This can result in misdiagnosis and suboptimal clinical decision-making. The heterogeneous expression of Ki-67 within tumours necessitates a more comprehensive assessment approach.

Radiomics is a new and exciting area that uses cutting-edge imaging methods to analyze tumour heterogeneity in great detail, offering a potential solution to these problems. It allows for quantitative feature extraction from popular imaging modalities like computed tomography (CT) and magnetic resonance imaging (MRI), which helps in understanding the morphology, texture, and intensity of tumours (7,8). As a critical biomarker linked with tumour growth and prognosis in different malignancies, Ki-67 expression has recently been the subject of an upsurge of research demonstrating the importance of radiomics in making such predictions (9). One study that used MRI-based radiomics signatures to predict Ki-67 expression before surgery in 179 patients with bladder cancer found a high area under the curve (AUC) of 0.859 on the training set and 0.819 on the validation set, suggesting good predictive performance and potential clinical value (10). Similarly, a study that looked at patients with NSCLC found that radiomics features taken from ¹⁸F-fluorodeoxyglucose positron emission tomography/computed tomography (F-FDG PET/CT) images could successfully differentiate Ki-67 status (11). The AUC values for the training and test datasets were 0.86 and 0.85, respectively. The results of the research highlight the promise of radiomics as a noninvasive method for predicting Ki-67 expression in various cancer types; this could lead to more targeted treatments and better patient outcomes.

The growing fascination with radiomics as a tool for predicting Ki-67 status highlights the promising possibilities of this technology to improve cancer diagnosis and therapy stratification (12). Advancements in methodology, including the use of a nomogram model to integrate radiomics characteristics with clinical factors, have the potential to enhance prediction efficiency and provide more tailored treatment options. However, existing radiomics studies for Ki-67 prediction in NSCLC mainly focus on intratumoural features, while ignoring the peritumoural region. Additionally, most models lack integration with clinical/CT features, which may limit their clinical utility as traditional CT signs have been associated with tumour proliferation. This study set out to a CT-based radiomics signature and create a nomogram that uses radscores in conjunction with clinical data to forecast the Ki-67 index in NSCLC patients. The primary objectives of this study are: (I) to develop a radiomics model using intratumoural, peritumoural (3 mm), and fused radiomics features; (II) to identify independent clinical and CT predictors for Ki-67 index; (III) to construct and internally/externally validate a combined prediction model (integrating radiomics score with clinical/CT predictors) for Ki-67 index in stage T1 NSCLC. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-aw-1868/rc).

Methods

Patient characteristics

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Ethics CommitteesCommittees of Zhejiang Cancer Hospital [No. IRB-2025-1159(IIT)], First Affiliated Hospital of Huzhou University (No. 2024KYLL085-01), and Sir Run Run Shaw Hospital, School of Medicine, Zhejiang University (No. 2025SYFYLS-2127). Individual consent for this retrospective analysis was waived. Figure 1 shows the whole workflow of the investigation.

Figure 1 Workflow of necessary steps in current study. AUC, area under the curve; DCA, decision curve analysis; GLCM, gray-level co-occurrence matrix; GLDM, gray-level dependence matrix; GLSZM, gray-level size zone matrix; ImageFusion3 mm, image fusion of the intratumour and peritumoural 3 mm; INTRA, intratumour; IntraPeri3 mm, intratumour and peritumour 3 mm; LASSO, least absolute shrinkage and selection operator; MLP, multilayer perceptron; Peri3 mm, peritumour 3 mm; ROC, receiver operating characteristic.

We used the 8th edition of the TNM classification at three tertiary hospitals from January 2020 to August 2024 to select patients who had surgical resection for clinical stage T1 (tumour size ≤3 cm) NSCLC (Figure 2). In center 1 (Zhejiang Cancer Hospital), 243 patients met the eligibility criteria, among whom 44 had high Ki-67 expression and 199 had low Ki-67 expression. From these patients with low Ki-67 expression, 128 were randomly selected as the training set. Another 75 cases from the other two centers served as the external validation set, among which 23 patients were high expression of Ki-67.

Figure 2 Flowchart of the patient selection. Center 1, Zhejiang Cancer Hospital; Center 2, First Affiliated Hospital of Huzhou University; Center 3, Sir Run Run Shaw Hospital, School of Medicine, Zhejiang University. CT, computed tomography.

Sample size was determined based on the ‘events per variable’ (EPV) principle for logistic regression models. Assuming 10 predictors (5 radiomics, 5 clinical/CT) and a minimum EPV of 10, at least 100 outcome events (high Ki-67 expression) were required. In the development cohort, 44 patients had high Ki-67 expression, and the external validation cohort had 23, totaling 67 events. To address potential attrition, we expanded the sample size to 247 (development + internal validation) and 75 (external validation), ensuring EPV ≥5 [a commonly accepted minimum for preliminary models; Harrell et al., 2015 (13), Regression Modeling Strategies].

According to the 8th edition of the tumor-node-metastasis (TNM) staging recommendations, the tumour diameter on preoperative CT imaging could not exceed 3 cm. CT imaging had to be done within one month before surgery, and pathological confirmation of invasive lung adenocarcinoma had to be present. All patients underwent surgical resection of lung tumours (lobectomy or wedge resection) after preoperative CT imaging and Ki-67 index assessment was performed via postoperative pathological examination. No neoadjuvant therapy (chemotherapy, radiotherapy) was administered before surgery, as this may affect tumour radiomics features and Ki-67 expression.

Inclusion criteria: (I) pathological confirmation of invasive via postoperative histopathology; (II) preoperative staging as T1 (tumour size ≤3 cm) based on 8th TNM classification (assessed via chest CT and whole-body PET-CT if suspected metastasis); (III) chest CT examination performed within 1 month before surgery; (IV) no neoadjuvant therapy (chemotherapy, radiotherapy) administered before surgery.

Exclusion criteria: (I) prior or current diagnosis of another malignant tumour (except non-melanoma skin cancer); (II) more than one nodule seen on preoperative CT scans; (III) evidence of distant metastasis (confirmed via chest/abdominal CT, brain MRI, and whole-body PET-CT); (IV) lack of thin-section CT (1.25 mm) images; (V) incomplete clinical or pathological data.

CT acquisition and interpretation

A 64-row 128-slice spiral CT from Siemens (Forchheim, Germany) was used to do the chest scan. Beginning at the thoracic opening, scan down to the level of the diaphragm. After taking a deep breath in, the patients were asked to lie down on their backs and hold their breath. Tube voltage was 120 kV, tube current was 120 mA, window width was 1,300–1,500 Hounsfield units (HU), window position was −600 to −700 HU, pitch was 1.0, and frame rotation time was 0.33 s/360 degrees. These were the scanning parameters. Using the lung technique, we rebuilt the lung window with a 1.25 mm reconstruction thickness and 1.25 mm layer spacing. Both the thickness and the spacing of the layers used to rebuild the mediastinal window were 5 mm.

Two thoracic radiologists, one with a decade of expertise and the other with five, evaluated conventional imaging characteristics for the purpose of diagnosing lung nodules. They worked separately, without access to clinical or pathological data, to evaluate the CT pictures. We gained consensus through conversation where there was a dispute regarding the findings. The evaluated imaging features included density (solid, part-solid, or ground-glass), diameter, lobulation sign, bubble sign, spiculation, pleural indentation, cystic lung cancer, peripheral ground-glass opacity (GGO), bronchus sign, vacuolation and margin.

Ki-67 proliferation index (PI) assessment

In order to assess the tumour cells’ proliferation, immunohistochemistry was performed using a mouse anti-human Ki-67 monoclonal antibody (clone: MIB-1, Dako, Denmark) within 3 days after surgical resection. A cell was considered positive if its nucleus was brown in color. We found three spots in the tissue samples where the number of positively stained cells was significantly higher than in the others. The average percentage of positive cells throughout these three locations was calculated, and 100 cells were randomly counted in each region under high magnification (×400) to determine the Ki-67 index.

As the proportion of cancer cells that stained positive, the Ki-67 PI was documented. If the Ki-67 PI is less than 14%, it is considered to have low expression, and if it is equal to or greater than 14%, it is considered to have strong expression (14). Pathologists who assessed Ki-67 expression were blinded to patients’ clinical information (e.g., gender, age), CT features (e.g., lobulation, density), and radiomics analysis results to avoid assessment bias.

Image acquisition

Region of interest (ROI) segmentation

Two seasoned radiologists manually annotated the training dataset using ITK-SNAP to demarcate the ROI in our study. A veteran radiologist with more than 20 years of expertise clarified any inconsistencies in their annotations. We built an artificial segmentation model upon this base of training data provided by this painstaking hand demarcation.

Peritumoural region dilation

The peritumoural region, which surrounds the tumour, is critical for various medical analyses. We methodically expanded the manually delineated ROI at radial intervals of 3 mm to systematically explore the influence of varying peritumoural distances on the predictive accuracy of our model.

Radiomics procedure

Feature extraction

In this study, the voxels of all CT images were resampled to 1×1×1 mm using the trilinear interpolation method; the density values were discretized via equal-width binning into 32 intervals with a bin width of 10 HU; intensity normalization was performed using the Z-score method (mean =0, standard deviation =1); and the scanner-related variations across multiple centers were calibrated by the ComBat method. There are three separate classes of handcrafted characteristics: geometry, intensity, and texture. In order to provide a spatial depiction of the tumour’s structure, geometric features include all of its three-dimensional shape parameters. Intensity characteristics reveal information about the tumour’s intrinsic variability or homogeneity by capturing the first-order statistical aspects of the voxel intensities. Texture features capture the subtle spatial interactions among voxels and outline the second-order and higher-order spatial distributions of the intensities. These features depict more complex patterns. The neighborhood gray-tone difference matrix (NGTDM), gray-level size zone matrix (GLSZM), gray-level run length matrix (GLRLM), and gray-level co-occurrence matrix (GLCM) are among the methods used to extract these texture properties. There are a total of 1,403 features that were handcrafted: 1,040 texture features, 340 intensity features, and 23 geometry features. Our radiomics feature analysis program uses PyRadiomics (http://pyradiomics.readthedocs.io) to extract all of the handmade features.

Feature selection

We began by testing how well the image attributes held up under stress. Because the contours of the tumour subregions are used in the feature computation, we made sure to choose image features that can withstand uncertainty in tumour segmentation. Here, feature resilience was determined using a combination of test-retest and interrater analyses. A test-retest study was conducted using 30 patients selected at random from the discovery dataset. In this analysis, one rater segmented the tumour subregions twice for each patient. The interrater analysis dataset comprised an additional 30 cases selected at random, with two raters independently segmenting the tumour subregions for each patient. Using the intraclass correlation coefficient (ICC), we evaluated the features retrieved from these subregions that were divided into numerous segments. We regarded features resilient against intra- and interrater uncertainty if their ICCs were 0.85 or higher.

We further decreased the data dimension by excluding features with strong correlations. The Spearman correlation coefficient is computed here for every pair of characteristics. Only one of the two feature pairs with associated coefficients ≥0.90 was kept. The last 150 picture features are chosen for their robustness, predictiveness, and lack of redundancy.

Signature building

Radiomics signature

By employing least absolute shrinkage and selection operator (LASSO) for stringent feature selection, we constructed a radiomics risk model via the machine learning algorithm multilayer perceptron (MLP). Comparative analyses were performed to gauge each model’s performance, and we investigated the benefits of feature fusion by integrating multimodal features, allowing us to explore the advantages of combining multiple imaging modalities to increase predictive accuracy.

Periradiomics signature [peritumour 3 mm (Peri3 mm)]

We applied the same rigorous feature selection process used for the IntraRadiomics Signature. The final model was established via the same array of machine learning algorithms, ensuring consistency and comparability of our approach to both intra- and peritumoural analyses.

Metrics

By building receiver operating characteristic (ROC) curves to measure its discriminative capabilities, we thoroughly tested our deep learning model’s diagnostic efficacy in the test cohort. The Hosmer-Lemeshow goodness-of-fit test was used in conjunction with calibration curves to assess the dependability of the model during calibration. To further comprehend the possible advantages in a clinical setting, decision curve analysis (DCA) was used to evaluate our predictive models’ clinical value.

Statistical analysis

Missing data were assessed for all predictors: carcinoembryonic antigen (CEA) had the highest missing rate (5.2%, 13/247), while other predictors (gender, age, CT features) had no missing data. Missing CEA values were handled via multiple imputation (5 imputed datasets) using predictive mean matching, with gender, age, and tumour diameter as auxiliary variables. Sensitivity analysis using complete-case analysis (n=234) was performed to confirm consistency of results. The clinical traits were confirmed to be normal using the Shapiro-Wilk test. The distribution features of continuous variables dictated whether they were tested using the t-test or the Mann-Whitney U test. Chi-square (χ²) tests were used to analyze categorical variables. Continuous predictors (age, tumour diameter, CEA) were processed as follows: age and diameter were normally distributed and used as continuous variables, while CEA was log-transformed due to right skewness. Categorical predictors were coded as follows: gender (0= female, 1= male), density [0= pure ground-glass opacity (pGGO), 1= part-solid, 2= solid], and binary features (e.g., lobulation sign: 0= absence of lobulation, 1= mild lobulation, 2= marked lobulation) as dummy variables. Table 1 details the baseline characteristics for all cohorts. Importantly, there were no significant differences between cohorts (P>0.05), proving that the groups were fairly divided.

Multivariate logistic regression was used for the clinical model and combined model; the radiomics model was built using a MLP machine learning algorithm.

Model-building steps for radiomics models: (I) feature extraction (1,403 handcrafted features via PyRadiomics); (II) feature filtering (ICC ≥0.85 for reliability, Spearman r<0.90 for no redundancy); (III) optimal feature selection (LASSO with 10-fold cross-validation); (IV) model training (MLP with 80% training/20% validation split within the development cohort).

Internal validation

10-fold cross-validation was performed on the development cohort to assess the stability of the radiomics model; bootstrapping (1,000 resamples) was used to validate the clinical and combined models.

External validation

For external validation, the combined model’s parameters (regression coefficients for gender, density, lobulation, and radiomics score) derived from the development cohort were directly applied to the external validation set to calculate individual predicted probabilities of high Ki-67 expression, without re-estimating model parameters (to assess transportability).

We used Python 3.7.12 on the OnekeyAI platform, version 4.9.1, to conduct all of our data analysis. We used Statsmodels version 0.13.2 to do our statistical analyses. The extraction of radiomics features was carried out using PyRadiomics version 3.0.1. Thanks to Scikit-learn version 1.0.2, the MLP machine learning implementation was made possible.

Results

Clinical characteristics

For the model development and evaluation, the training set comprised 172 participants, of whom 44 had high Ki-67 expression (the outcome event). The internal test set comprised 75 participants, with 23 cases of high Ki-67 expression. Comparison between development and external validation cohorts: (I) setting: all centers were tertiary hospitals, but the external validation cohort had a higher proportion of rural patients (35.2% vs. 22.1% in development cohort). (II) Eligibility criteria: identical for both cohorts. (III) Outcome: same Ki-67 assessment method (MIB-1 antibody, 14% cutoff). (IV) Predictors: CEA missing rate was higher in the external validation cohort (7.1% vs. 5.2%), but other predictors had no missing data. No significant differences in baseline characteristics (gender, age, tumour diameter, CT features) were observed between cohorts (all P>0.05, Table 1). To assess the unadjusted associations between candidate clinical predictors and the Ki-67 index status, univariable logistic regression analyses were performed. The results are detailed in Table 2. Univariate logistic regression showed that 12 factors were significantly associated with high Ki-67 expression (all P<0.05): gender [male vs. female, odds ratio (OR) =0.178, P<0.001], age (per 1-year increase, OR =0.984, P<0.001), tumour diameter (per 1-cm increase, OR =0.65, P<0.001), lobulation sign (present vs. absent, OR =0.481, P<0.001), bubble sign (OR =0.329, P<0.001), spiculation (OR =0.367, P<0.001), pleural indentation (OR =0.537, P=0.018), peripheral GGO (OR =0.162, P<0.001), bronchus sign (OR =0.250, P<0.001), vacuolation (OR =0.291, P<0.001), margin (OR =0.344, P<0.001), and density (solid vs. non-solid, OR =0.491, P<0.001). Furthermore, a multivariate logistic regression analysis was conducted with the aforementioned indicators. There were three independent characteristics that were found to differentiate low or high expression of the Ki-67 index: gender (P=0.006), the Lobulation sign (P=0.03), and density (P=0.03). After that, we used those independent factors to build a clinical model.

Radiomics feature screening and model construction

After LASSO, 15, 14, 12 and 16 optimal radiomics features were finally screened out via the intratumour and peritumour 3 mm (IntraPeri3 mm), intratumour (INTRA), image fusion of the intratumour and peritumoural 3 mm (ImageFusion3 mm) and Peri3 mm methods, respectively (Figure 2). Then, the IntraPeri3 mm, INTRA, ImageFusion3 mm and Peri3 mm models were constructed with these features. The AUC of the IntraPeri3 mm model 0.891 [95% confidence interval (CI): 0.8388–0.9439] for predicting Ki-67 was greater than that of the INTRA model 0.856 (95% CI: 0.7975–0.9140) in the training set (P=0.02). The AUC of the IntraPeri3 mm model 0.880 (95% CI: 0.7956–0.9640) was greater than that of the Peri3 mm model 0.743 (95% CI: 0.6202–0.8649) in the test set (P=0.009). Figures 3,4 show that the IntraPeri3 mm model had slightly higher AUC 0.891 (95% CI: 0.8388–0.9439) for Ki-67 prediction in the training set compared to the ImageFusion3 mm model 0.869 (95% CI: 0.8105–0.9268), but these differences were not statistically significant (all P>0.05). The best radiomics model was considered to be the IntraPeri3 mm model (Table 3).

Figure 3 ROC curves of the combined model, clinical models, and radiomics model were compared in the training (A,B) and testing (C,D) cohorts to predict Ki-67 index status. AUC, area under the curve; CI, confidence interval; ImageFusion3 mm, image fusion of the intratumour and peritumoural 3 mm; INTRA, intratumour; IntraPeri3 mm, intratumour and peritumour 3 mm; Peri3 mm, peritumour 3 mm; ROC, receiver operating characteristic.

Figure 4 DeLong test of the combined model, clinical models, and radiomics model in the training (A,C) and testing (B,D) cohorts. ImageFusion3 mm, image fusion of the intratumour and peritumoural 3 mm; INTRA, intratumour; IntraPeri3 mm, intratumour and peritumour 3 mm; Peri3 mm, peritumour 3 mm.

Construction of the combined model

A combined model was built using the Radscore of the IntraPeri3 mm model and the separate clinical and CT predictors; Figure 5 shows the nomogram of this model. For the training set, the AUC for the combined model was 0.904 vs. 0.778 (P<0.001), and for the IntraPeri3 mm model, it was 0.891 vs. 0.778 (P<0.01), while for the test set, the AUC for the combined model was 0.882 vs. 0.775 (P<0.05) compared to the AUC for the clinical model. Check out Figures 3,4 and Table 3.

Figure 5 The nomogram is developed to predict the risk of Ki-67 index status, which combining of Rad-score, gender, density and lobulation. Gender: 0= female, 1= male; density: 0= pure ground-glass opacity, 1= part-solid, 2= solid; lobulation sign: 0= absence of lobulation, 1= mild lobulation, 2= marked lobulation.

The full specification of the combined model is presented to allow for individual prediction. The model was defined by the following logistic regression equation: Logit(P) = −1.235 + (0.876 × X_gender) + (1.521 × X_density) + (1.392 × X_lobulation) + (2.105 × Radscore), P=1/[1 + e^(−Logit(P)].

To use the combined model for individual prediction: (I) for a given patient, assign points for each predictor based on Figure 5: e.g., male (10 points), solid density (20 points), lobulation sign present (15 points), IntraPeri3 mm radiomics score=0.8 (30 points). (II) Sum the points (10+20+15+30=75 points). (III) Locate the total points on the ‘Total Points’ axis and draw a vertical line to the ‘Risk’ axis to obtain the predicted probability of high Ki-67 expression (e.g., 75 points correspond to a 65% risk). Alternatively, use the regression equation: logit(P) = −1.235 + 0.876 × gender (1= male, 0= female) + 1.521 × density (2= solid, 1= part-solid, 0= pGGO) + 1.392 × lobulation (1= present, 0= absent) + 2.105 × radiomics score, where P is the probability of high Ki-67 expression.

The clinical model, Model GPT, and the combined model all had reasonably high degrees of calibration, as shown by the calibration curves (Figure 6). The combined model showed good net clinical benefit when the threshold was set between 0.10 and 0.80, according to DCA. Make note of Figure 7.

Figure 6 Calibration curves were plotted for the training (A) and testing cohorts (B). IntraPeri3 mm, intratumour and peritumour 3 mm.

Figure 7 DCA was performed for each model in the training and testing cohorts. DCA, decision curve analysis; IntraPeri3 mm, intratumour and peritumour 3 mm.

Discussion

The most prevalent type of lung cancer is NSCLC, and there is a strong correlation between tumour growth and prognosis and the expression level of the biomarker Ki-67 (15-17). Radiomics is a relatively new field of study that uses high-dimensional analysis of medical imaging data to potentially uncover biological information, which in turn helps us understand tumour features better (18-20). The purpose of this research was to find a way to determine if a patient with advanced NSCLC had a positive or negative Ki-67 score without invasive procedures. As far as we know, this is the first study to build a combined model that predicts the Ki-67 index status of patients with NSCLC by merging tumour and peritumoural radiomics methods with clinical and CT data. The prediction accuracy of Ki-67 can be enhanced through the combination of radiomics with traditional prognostic indicators. This, in turn, can lead to better customized treatment solutions for patients with NSCLC.

Higher Ki-67 levels usually indicate a higher tumour grade (21,22), and the link between tumour grade and Ki-67 expression is well-established. Extensive research has linked elevated Ki-67 expression to poor differentiation and higher histological grades in a number of cancer types, including gastric and breast malignancies. As an example, a meta-analysis found that gastric cancer patients with overexpressed Ki-67 had more advanced tumour stages and a worse overall prognosis (23). The luminal A and B subtypes of breast cancer can be distinguished by Ki-67, with greater levels indicating more aggressive disease (24). Furthermore, Ki-67 levels were observed to correlate significantly with histological grade, further supporting its function as a proliferation marker that represents tumour biology (25). Ki-67 is a great tumour grade indicator because it helps us understand how cancers work biologically and makes it easier to decide how to treat them (26). Critical for diagnosing the malignancy of pulmonary diseases, CT scans show a strong correlation between tumour density and the degree of differentiation of lung cancer (27-29). A correlation between aggressive tumour behavior and the density of these nodules on imaging investigations was shown by the dramatically increased proportion of high Ki-67 expression in solid lung nodules, which is a hallmark of cellular proliferation. The lobulation sign, a common CT sign of lung cancer, was also found to be an independent predictor of Ki-67 expression in this study, which differs from previous research. The degree to which lung cancer has spread to the periphery is often indicated by the lobulation sign (30,31). Several malignancies express Ki-67 differently in male and female patients, according to the research (32). Ki-67 is an important marker of cellular proliferation. In the case of lung adenocarcinoma, for instance, researchers found that Ki-67 expression was greater in male patients than in female patients, especially in instances of tumours with weak differentiation and solid subtypes. In male patients, who often exhibit more severe manifestations of the disease, these results imply that Ki-67 could be a useful prognostic marker (3). Such results are in line with the study’s conclusion that sex is a separate determinant of elevated Ki-67 expression. In addition, univariate analysis showed that most malignant imaging features were associated with a reduced risk of adverse outcomes in this study. This counterintuitive result suggests the presence of strong confounding factors. After adjustment by multivariate logistic regression, two independent risk factors were identified: solid component of the tumor and lobulation sign. Other features, such as spiculation and pleural indentation, did not show independent predictive value. These findings indicate that when assessing the risk of pulmonary nodules/lung cancer, emphasis should be placed on the comprehensive evaluation of the solid component and lobulation sign. For lesions with these features, even if other morphological characteristics are atypical, their adverse biological behaviors should be vigilant.

Tumour size, form, and edge features are examples of traditional features that are chosen and retrieved by hand. In most cases, radiologists’ expertise is required for these qualities. Advanced features, on the other hand, are those that deep learning and machine learning algorithms automatically extract. Common examples of such features are texture features and GLCM features, both of which are capable of capturing more intricate information about images. Better and more precise disease information may be retrieved with the help of sophisticated features, which can sift through mountains of data in search of any patterns and relationships. For instance, when it comes to forecasting how NSCLC patients will react to radiation, several studies have demonstrated that the advanced features retrieved from radiomics work better than conventional features (33). Particularly in investigations including Ki-67 expression, radiomics characteristics have demonstrated promise in foretelling tumour biological behavior. The predictive role of radiomics in predicting Ki-67 expression in lung cancer has recently been the subject of a few articles (34). However, these studies are limited to feature extraction of the tumour itself and do not include the peritumoural area. There is a 2–5 mm transition zone around lung adenocarcinoma, which contains information reflecting the invasive biological behavior of lung cancer (35). Radiomics can reflect tumour heterogeneity. In this study, an outward expansion of 3 mm from the INTRA was taken as the peritumoural area to include the transition zone as much as possible. To extract features and conduct better analyses, this study also adopted image prefusion (ImageFusion3 mm) and feature postfusion (IntraPeri3 mm) methods for feature extraction and modeling of the tumour and the peritumoural area, respectively. For the most part, the INTRA model is first-order. Using standard and elementary metrics, first-order statistics illustrate the distribution of voxel intensities inside the mask-defined image region. For NSCLC, higher CT scan values related to first-order radiomics features often indicate a higher level of malignancy (36). Compared with the manual measurement of CT values in CT images, radiomics is more accurate and offers a more comprehensive analysis. In the Peri3 mm model, the main features are the radiomics features of the GLCM. Owing to the existence of the transition zone in the surrounding area of lung cancer, analyzing this area usually enables us to obtain relatively effective radiomics features. In the imaging diagnosis of the transition zone, GLCM features can effectively describe the texture of pulmonary nodules. The texture of pulmonary nodules is an important indicator for determining whether they are benign or malignant. Through the contrast feature of the GLCM, the differences in this texture can be quantified. Nodules with high contrast have relatively large differences in pixel gray values and may appear as uneven internal structures of the nodules on the images, which may indicate a greater degree of malignancy. In the postfusion model, the above two types of features are adopted. However, in the prefusion model, neither of these two types of features is prominent. The result is that the postfusion model provides better predictions than the prefusion model.

The combined model’s AUC in the external validation set (0.865) was slightly lower than that in the development set (0.904), but remained higher than existing radiomics models for Ki-67 prediction in NSCLC [e.g., Fu et al., 2021 (34), AUC =0.82; Liu et al., 2023 (9), AUC =0.84]. This suggests that integrating peritumoural radiomics with clinical/CT features improves discriminative performance, and the model is relatively transportable across centers. Overall, this study achieved its primary objectives: we developed a multivariable model integrating intratumoural/peritumoural radiomics with clinical/CT features and validated it in an independent cohort. The model showed good discrimination (AUC >0.85) and calibration (Hosmer-Lemeshow P>0.05) across datasets, addressing the limitations of existing models that ignore peritumoural regions.

This investigation is not without its constraints. First, the limited sample size necessitates further validation through studies involving larger populations. Bigger sample sizes from multicenter studies should be the focus of future research efforts. Secondly, no uncommon pathological kinds were included in this study; instead, it only included adenocarcinoma subtypes of NSCLC. Lastly, a variety of instruments were used to gather the samples. Despite uniform processing of the images, the specific equipment used may still influence the findings. Finally, the Ki-67 index status in NSCLC patients can be accurately predicted by combining intratumoural and peritumoural radiomics features with clinical and CT features.

Conclusions

This study develops and validates a non-invasive combined model for predicting the Ki-67 index in stage T1 NSCLC patients, integrating intratumoural/peritumoural 3 mm radiomics features (with the IntraPeri3 mm model outperforming standalone regional models) and independent clinical (gender) and CT (density, lobulation) predictors; the combined model achieves excellent discriminative performance (AUC =0.904 in the training set and 0.882 in the external validation set), good calibration, and significant net clinical benefit across a threshold range of 0.10–0.80, addressing existing radiomics research limitations by highlighting peritumoural region significance and multi-modal fusion, providing a user-friendly nomogram for preoperative quantitative prediction of Ki-67 expression to enable early risk stratification, reduce reliance on invasive biopsies, avoid tumour heterogeneity-induced misdiagnosis, and guide personalized treatment strategies, while future studies should expand sample size, include diverse NSCLC pathological subtypes, validate in multi-ethnic populations, and integrate artificial intelligence (AI)-based automatic segmentation and dynamic imaging features to enhance generalizability, efficiency, and predictive accuracy for routine clinical translation.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-aw-1868/rc

Data Sharing Statement: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-aw-1868/dss

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-aw-1868/prf

Funding: This work was supported by a grant from Zhejiang Provincial Traditional Chinese Medicine Science and Technology Project (No. 2024ZL451).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-aw-1868/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Ethics Committees of Zhejiang Cancer Hospital [No. IRB-2025-1159(IIT)], First Affiliated Hospital of Huzhou University (No. 2024KYLL085-01), and Sir Run Run Shaw Hospital, School of Medicine, Zhejiang University (No. 2025SYFYLS-2127). Individual consent for this retrospective analysis was waived.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Ni B, Kong X, Yan Y, et al. Combined analysis of gut microbiome and serum metabolomics reveals novel biomarkers in patients with early-stage non-small cell lung cancer. Front Cell Infect Microbiol 2023;13:1091825. [Crossref] [PubMed]
2. Roi C, Riviș M, Roi A, et al. CD34 and Ki-67 Immunoexpression in Periapical Granulomas: Implications for Angiogenesis and Cellular Proliferation. Diagnostics (Basel) 2024;14:2446. [Crossref] [PubMed]
3. Ma X, Zhou S, Huang L, et al. Assessment of relationships among clinicopathological characteristics, morphological computer tomography features, and tumor cell proliferation in stage I lung adenocarcinoma. J Thorac Dis 2021;13:2844-57. [Crossref] [PubMed]
4. Al-Keilani MS, Elstaty R, Alqudah MA. The Prognostic Potential of Neurokinin 1 Receptor in Breast Cancer and Its Relationship with Ki-67 Index. Int J Breast Cancer 2022;2022:4987912. [Crossref] [PubMed]
5. Luo X, Zheng R, Zhang J, et al. CT-based radiomics for predicting Ki-67 expression in lung cancer: a systematic review and meta-analysis. Front Oncol 2024;14:1329801. [Crossref] [PubMed]
6. Pelosi G, Travis WD. Head-to-head: Should Ki67 proliferation index be included in the formal classification of pulmonary neuroendocrine neoplasms? Histopathology 2024;85:535-48. [Crossref] [PubMed]
7. Shang J, Guo Y, Ma Y, et al. Cardiac computed tomography radiomics: a narrative review of current status and future directions. Quant Imaging Med Surg 2022;12:3436-53. [Crossref] [PubMed]
8. Ger RB, Wei L, Naqa IE, et al. The Promise and Future of Radiomics for Personalized Radiotherapy Dosing and Adaptation. Semin Radiat Oncol 2023;33:252-61. [Crossref] [PubMed]
9. Liu F, Li Q, Xiang Z, et al. CT radiomics model for predicting the Ki-67 proliferation index of pure-solid non-small cell lung cancer: a multicenter study. Front Oncol 2023;13:1175010. [Crossref] [PubMed]
10. Zheng Z, Gu Z, Xu F, et al. Magnetic resonance imaging-based radiomics signature for preoperative prediction of Ki67 expression in bladder cancer. Cancer Imaging 2021;21:65. [Crossref] [PubMed]
11. Manafi-Farid R, Askari E, Shiri I, et al. [18F]FDG-PET/CT Radiomics and Artificial Intelligence in Lung Cancer: Technical Aspects and Potential Clinical Applications. Semin Nucl Med 2022;52:759-80.
12. Sun H, Zhou P, Chen G, et al. Radiomics nomogram for the prediction of Ki-67 index in advanced non-small cell lung cancer based on dual-phase enhanced computed tomography. J Cancer Res Clin Oncol 2023;149:9301-15. [Crossref] [PubMed]
13. Harrell FE Jr. Multivariable modeling strategies. In: Regression modeling strategies: With applications to linear models, logistic and ordinal regression, and survival analysis. Cham: Springer International Publishing, 2015:63-102.
14. Zhang L, Du Q, Shen M, et al. Interpretable model based on MRI radiomics to predict the expression of Ki-67 in breast cancer. Sci Rep 2025;15:13318. [Crossref] [PubMed]
15. De Zuani M, Xue H, Park JS, et al. Single-cell and spatial transcriptomics analysis of non-small cell lung cancer. Nat Commun 2024;15:4388. [Crossref] [PubMed]
16. Yan Y, Sun D, Hu J, et al. Multi-omic profiling highlights factors associated with resistance to immuno-chemotherapy in non-small-cell lung cancer. Nat Genet 2025;57:126-39. [Crossref] [PubMed]
17. Zhang WQ, Gao JN, Chen XL, et al. PPM1H is an independent prognostic biomarker of non-small cell lung cancer. Neoplasma 2021;68:917-23. [Crossref] [PubMed]
18. Molinari V, Gitto S, Serpi F, et al. Radiomics for bone tumour diagnosis and management. Clin Radiol 2025;91:107062. [Crossref] [PubMed]
19. McCague C, Ramlee S, Reinius M, et al. Introduction to radiomics for a clinical audience. Clin Radiol 2023;78:83-98. [Crossref] [PubMed]
20. Temperley HC, O'Sullivan NJ, Waters C, et al. Radiomics; Contemporary Applications in the Management of Anal Cancer; A Systematic Review. Am Surg 2024;90:445-54. [Crossref] [PubMed]
21. Wang J, Gao W, Lu M, et al. Development of an interpretable machine learning model for Ki-67 prediction in breast cancer using intratumoral and peritumoral ultrasound radiomics features. Front Oncol 2023;13:1290313. [Crossref] [PubMed]
22. Yang H, Wang W, Cheng Z, et al. Radiomic Machine Learning in Invasive Ductal Breast Cancer: Prediction of Ki-67 Expression Level Based on Radiomics of DCE-MRI. Technol Cancer Res Treat 2024;23:15330338241288751. [Crossref] [PubMed]
23. Zhao P, Li Y, Lu Y. Aberrant expression of CD133 protein correlates with Ki-67 expression and is a prognostic marker in gastric adenocarcinoma. BMC Cancer 2010;10:218. [Crossref] [PubMed]
24. Mushtaq M, Chaudry SS, Khalid Sheikh A, et al. Comparison of Different Molecular Subtypes with 14% Ki-67 Cut-off Threshold in Breast Cancer Patients of Pakistan- An Indication of Poor Prognosis. Arch Iran Med 2021;24:837-44. [Crossref] [PubMed]
25. Kubik S, Jesionek-Kupnicka D, Stawiski K, et al. Clinical significance of CD34, podoplanin and Ki-67 expression in patients with locally advanced squamous cell cervical carcinoma. J Obstet Gynaecol 2022;42:2248-54. [Crossref] [PubMed]
26. Galeano B, Smith CJ, Yi ES, et al. Ki-67 Proliferation Index Is Associated With Tumor Grade and Survival in Pleural Epithelioid Mesotheliomas. Am J Surg Pathol 2024;48:615-22. [Crossref] [PubMed]
27. Wen LJ, Zhao QY, Yin YH, et al. Application value of double-layer spectral detector CT in differentiating central lung cancer from atelectasis. Ann Palliat Med 2022;11:1990-6. [Crossref] [PubMed]
28. Sekine Y, Sugai K, Kuroda K, et al. Investigation of the integration of computed tomography (CT) value doubling time for lung cancer with subsolid nodules using a three-dimensional image analysis system. Clin Radiol 2025;83:106813. [Crossref] [PubMed]
29. Li Q, Li X, Li XY, et al. Spectral CT in Lung Cancer: Usefulness of Iodine Concentration for Evaluation of Tumor Angiogenesis and Prognosis. AJR Am J Roentgenol 2020;215:595-602. [Crossref] [PubMed]
30. Wang Q, Fu C, Wang K, et al. Clinical Multi-features Analysis of Cystic Lung Adenocarcinoma and Construction of Invasive Risk Prediction Model. Zhongguo Fei Ai Za Zhi 2024;27:266-275. [Crossref] [PubMed]
31. Pan X, Wang H, Yu H, et al. Lung cancer associated with cystic airspaces: CT and pathological features. Transl Cancer Res 2020;9:3960-4. [Crossref] [PubMed]
32. Fainstein Day P, Glerean M, Lovazzano S, et al. Gender differences in macroprolactinomas: study of clinical features, outcome of patients and ki-67 expression in tumor tissue. Front Horm Res 2010;38:50-8. [Crossref] [PubMed]
33. Kawahara D, Imano N, Nishioka R, et al. Prediction of radiation pneumonitis after definitive radiotherapy for locally advanced non-small cell lung cancer using multi-region radiomics analysis. Sci Rep 2021;11:16232. [Crossref] [PubMed]
34. Fu Q, Liu SL, Hao DP, et al. CT Radiomics Model for Predicting the Ki-67 Index of Lung Cancer: An Exploratory Study. Front Oncol 2021;11:743490. [Crossref] [PubMed]
35. Liu X, Guo R, Li D, et al. Homotypic cell-in-cell structure as a novel prognostic predictor in non-small cell lung cancer and frequently localized at the invasive front. Sci Rep 2024;14:18952. [Crossref] [PubMed]
36. Gao S, Wu Y, Li H, et al. The predictive value of computed tomography value on high-resolution images in differentiating invasive from indolent lung adenocarcinoma. Transl Cancer Res 2024;13:268-77. [Crossref] [PubMed]