Radiomics differences between GOLD I–II and GOLD III–IV in patients with chronic obstructive pulmonary disease

Abstract

Background

High-resolution computed tomography (HRCT) serves as an effective imaging modality for characterizing and quantifying structural lung changes associated with chronic obstructive pulmonary disease (COPD). The emerging field of radiomics enables the rapid extraction of numerous quantitative features from medical images, offering considerable promise for supporting clinical decisions. Therefore, the aim of this study was to explore the lung function-associated radiomics features in COPD patients.

Methods

This cross-sectional study enrolled 223 patients diagnosed with COPD. The final study cohort comprised 94 patients classified as Global Initiative for Chronic Obstructive Lung Disease (GOLD) stage I–II and 56 as GOLD stage III–IV. For all participants, baseline demographic and clinical characteristics, spirometry data, and chest HRCT images were collected. Subsequently, 944 quantitative radiomics features were extracted from each HRCT scan. To identify features associated with GOLD stages, we employed least absolute shrinkage and selection operator (LASSO) regression for feature selection, followed by logistic regression for modeling. Finally, a nomogram along with its validation curves was constructed to visualize and assess the performance of the predictive model.

Results

Following feature selection via LASSO regression, eight potential predictors were retained. Subsequent binary logistic regression refined this set, revealing two radiomics features (original_firstorder_10Percentile and wavelet.LHL_glszm_GrayLevelVariance) that were independently associated with GOLD stages. The model’s discrimination was validated by a C-index of 0.838 and an area under the receiver operating characteristic curve (AUC) of 0.820. Furthermore, decision curve analysis (DCA) confirmed the clinical utility of the nomogram, showing a positive net benefit for decision thresholds from 0.13 to 0.84.

Conclusions

Collectively, a noticeable difference in radiomics features was observed between GOLD I–II and GOLD III–IV patients with COPD. The computed tomography (CT)-based radiomics features, original_firstorder_10Percentile and wavelet.LHL_glszm_GrayLevelVariance, can be potentially used to evaluate the severity of COPD patients. However, our results require further validation through multicenter and large-scale clinical studies.

Highlight box.

Key findings

• Computed tomography-based radiomics features, original_firstorder_10Percentile and wavelet.LHL_glszm_GrayLevelVariance, can be potentially used to predict the severity of chronic obstructive pulmonary disease (COPD).

What is known and what is new?

• COPD staging relies heavily on spirometry, which suffers from limited accessibility in resource-constrained settings and operator-dependent variability.

• Our work establishes high-resolution computed tomography radiomics as a complementary tool for objectively stratifying COPD severity, addressing an urgent need for scalable diagnostic methods.

What is the implication, and what should change now?

• This study is the first to report specific radiomic signatures distinguishing mild-to-moderate from severe COPD. The features original_firstorder_10Percentile and wavelet.LHL_glszm_GrayLevelVariance quantify subtle parenchymal alterations invisible to human observers, offering a pathway for automated severity classification.

Introduction

Chronic obstructive pulmonary disease (COPD), the most common chronic respiratory disease, is characterized by irreversible airflow limitation, emphysema, and extrapulmonary manifestations, with high morbidity and mortality (1,2). Pulmonary function test (PFT), particularly the forced expiratory volume in 1 second per forced vital capacity (FEV1/FVC) and FEV1 percentage of predicted (FEV1% predicted), serves as the primary criterion for the diagnosing and severity evaluation of COPD (3). However, the utilization of PFT in China remains insufficient (4). The accuracy of PFT is influenced by various factors, such as the skill of the operator, patient cooperation, and equipment calibration, etc. (5-7). Furthermore, the specific location, extent, and subtle manifestations of the disease cannot be visually depicted, which hinders early diagnosis and the assessment of treatment (8).

In contrast, high-resolution computed tomography (HRCT) is a common and conventional medical examination in the diagnosis and assessment of airways diseases, particularly COPD. Sibtain et al. found that the alterations in lung parenchyma, small airways, and pulmonary vasculature, which were revealed by chest HRCT, provided superior anatomical resolution during both functional decline and the aging process (9). As artificial intelligence (AI) applications in HRCT imaging advance rapidly, HRCT scan shows a significant potential value in COPD diagnosis and management, even if it is not being widely utilized for the quantitative evaluation of COPD in current clinical practice. Meanwhile, the assessment of lung parenchymal and small airway abnormalities remains largely dependent on the experience of radiologists (10). As the prevalence of COPD continues rising, the visual evaluation of lung lesions by radiologists becomes increasingly burdensome (11).

Radiomics, representing an innovative technology, was initially introduced by Philippe Lambin of the Netherlands in 2012 (12). This methodology bridges medical imaging with precision medicine through the high-throughput extraction of quantitative features, converting medical images into high-dimensional datasets for subsequent data analysis aimed at addressing clinical challenges. Radiomics has the capability to generate hundreds to thousands of radiomic features, broadly categorized into first-order statistics, textural features, and high-order filter transform features. These features are instrumental in identifying and assessing specific diseases. The analysis of image texture encompasses the spatial relationships between varying intensity levels within the image, providing richer information than densitometry, thereby facilitating more precise diagnoses during the early stages of lesion development (13,14).

Recently, radiomics has gained wide application in oncology research (15,16), however, its utilization in the clinical diagnosis and evaluation of non-neoplastic diseases is still limited. Given that COPD is a disease with apparent lung structure pathological alterations and high heterogeneity, chest computed tomography (CT) image texture analysis potentially serves as a more effective method for quantifying pulmonary structure abnormality. Therefore, the purpose of this study was to explore the radiomics differences between mild-to-moderate and severe-to-very severe COPD patients, and to develop a radiomics model to predict COPD severity. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1-2721/rc).

Methods

Study design and population

This prospective cross-sectional study was carried out in the Respiratory Medicine and Critical Care Medicine of Suining Central Hospital between January 2023 and August 2024. This study was approved by the Research Ethics Committees of the Suining Central Hospital (No. KYLLKS20230103), in accordance with the Declaration of Helsinki and its subsequent amendments. Individual consent for this retrospective analysis was waived.

Sample size determination

Based on the previous studies (17,18), it was reported that the ratio of Global Initiative for Chronic Obstructive Lung Disease (GOLD) I–II to GOLD III–IV in COPD patients was around 2:1. Based on our earlier work (19-21), we calculated the required sample size. Briefly, a minimum of 144 participants (96 GOLD I–II participants and 48 GOLD III–IV participants) were required with effect size =0.4, power =0.8, α=0.05, and allocation ratio =2:1.

Inclusion and exclusion criteria

The inclusion criterion was stable COPD patients (age ≥40 years) with a post-bronchodilator ratio of FEV1/FVC <0.70. Exclusion criteria were as follows: patients without lung function test, history of thoracic surgery, other lung diseases [asthma, bronchiectasis, pleural effusion, pneumoconiosis, active pulmonary tuberculosis (TB), interstitial lung diseases (ILDs), pneumonia, pulmonary thromboembolism (PTE), pulmonary edema, chronic cor pulmonale (CCP)], pulmonary nodules classified as Lung Imaging Reporting and Data System (Lung-RADS) grade ≥3, connective tissue diseases (CTD), heart failure, renal failure, liver failure, and history of malignant diseases. A total of 223 COPD patients were recruited. In the end, 94 GOLD I–II patients and 56 GOLD III–IV patients were included (Figure 1).

Figure 1.

Summary of study design and data analysis. CIC, clinical impact curve; COPD, chronic obstructive pulmonary disease; DCA, decision curve analysis; GOLD, Global Initiative for Chronic Obstructive Lung Disease; LASSO, least absolute shrinkage and selection operator; LU-RADS, Lung Imaging Reporting and Data System; ROC, receiver operating characteristic.

Definitions

According to GOLD guidelines, the diagnosis of COPD was confirmed by the pulmonologists. The severity of lung function impairment is classified as follows: GOLD I (mild, FEV1 ≥80% of the predicted value), GOLD II (moderate, 50%≤ FEV1 <80% of the predicted value), GOLD III (severe, 30%≤ FEV1 <50% of the predicted value), and GOLD IV (very severe, FEV1 <30% of the predicted value). Individuals were classified as ex-smokers if they had ceased smoking for a period of 6 months or longer.

Clinical data and chest HRCT image collection

Demographic data, underlying diseases, and spirometry were recorded and collected. Meanwhile, the images were acquired using a 64-slice or higher-resolution CT scanner, with a reconstructed slice thickness of 0.625 mm. Prior to the imaging procedure, all participants underwent breath-hold training to ensure the reproducibility and consistency of the scans. The resulting images, extending from the lung apex to the lung base, were free of respiratory motion artifacts.

Volume of interest (VOI) segmentation

According to the previous study (22), the pulmonary radiomics data were collected. Briefly, the CT scan data, stored in DICOM format, were uploaded to the open-source software package 3D Slicer version 5.40 (https://www.slicer.org). For each case, we systematically selected 42 non-overlapping regions of interest (ROIs) from 11 axial CT slices. Each ROI was a circle with a diameter of 20 mm, ensuring uniformity in size. During delineation, care was taken to avoid mediastinal structures, fissures, dilated bronchi, and bullae. No distinction or avoidance was made for small airways (diameter <2 mm) or between panlobular and centrilobular emphysema patterns. The ROIs spanned from the apex to the base of the lung, encompassing 11 key anatomical landmarks: the sternoclavicular joint, aortic arch, azygous vein arch, right upper lobe bronchus, left upper lobe bronchus, left lower dorsal segment bronchus, right middle lobe bronchus, left lower pulmonary vein trunk, basal trunk bronchus, right lower pulmonary vein trunk, and diaphragmatic dome. Two ROIs were delineated at the level of the sternoclavicular joint, while four ROIs were delineated at each of the remaining 10 levels. Consequently, a total of 42 ROIs were obtained from the CT images of each patient.

Radiomics workflow

The workflow of radiomics encompasses image acquisition, ROI segmentation, radiomics feature extraction, feature selection, model construction in the training cohorts, and evaluation of the radiomics models’ performance (Figure 2).

Figure 2.

Radiomics workflow. COPD, chronic obstructive pulmonary disease; HRCT, high-resolution computed tomography.

Feature extraction, selection, and model construction

A total of 944 radiomics features were extracted from the chest HRCT images of each participant after resampling the images to a resolution of 1 mm × 1 mm × 1 mm using 3D Slicer version 5.40. These features comprised 32 first-order features, 168 texture features, and 744 high-level features derived from wavelet transformation. The features were standardized via Z-score normalization to eliminate discrepancies in numerical scale.

Statistical analysis

R version 4.3.1 was utilized for data processing and statistical analysis. Continuous variables were tested for distribution with the Kolmogorov-Smirnov test. Based on the test results, they were expressed accordingly: those following a normal distribution as mean ± standard deviation (SD), and those with a non-normal distribution as median and interquartile range [median (Q1, Q3)]. Categorical variables were expressed in frequencies. For comparisons, the Mann-Whitney U test was applied to continuous variables with non-normal distribution and to ordinal variables, while the Chi-square test was used for categorical variables. Multicollinearity among variables was evaluated by calculating the variance inflation factor (VIF). Variables exhibiting high multicollinearity (VIF ≥10) were removed from further analysis. Potential radiomics features associated with GOLD stages were screened via least absolute shrinkage and selection operator (LASSO) regression. A binary logistic regression model was then built using the features selected by LASSO to investigate independent associations with GOLD stages in COPD patients. This final model was validated and graphically represented using a nomogram, calibration curve, receiver operating characteristic (ROC) curve, decision curve analysis (DCA), and clinical impact curve (CIC) (Figure 1). A threshold of P<0.05 was set for statistical significance.

Results

Demographic data

In our study, 150 patients with stable COPD were finally included. Among them, 94 patients (62.6%) were GOLD I–II stages, and 56 patients (37.6%) were GOLD III–IV stages. Compared to GOLD I–II group, the patients in GOLD III–IV group had lower body mass index (BMI) and poorer lung function, encompassing FEV1% predicted, FEV1/FVC, and maximal mid-expiratory flow (MMEF) (all P<0.05) (Table 1).

Table 1. Demographic data of patients with COPD (n=150).

Variable	GOLD I–II (n=94)	GOLD III–IV (n=56)	Statistical value	P value
Age (years)†	70 [61, 73.75]	69 [62, 72]	−0.333	0.74
Sex (male)	70 (74.5)	49 (87.5)	3.635	0.057
BMI, kg/m2	24.13±3.5	21.94±2.82	3.434	<0.001
Smoking			−1.561	0.12
Non-smoking	35 (37.2)	13 (23.2)
Ex-smoking	32 (34.1)	23 (41.1)
Current-smoking	27 (28.7)	20 (35.7)
Underlying diseases/co-morbidities
Hypertension	23 (24.5)	11 (19.6)	0.466	0.50
T2DM	3 (3.2)	3 (5.4)	0.429	0.51
CAD	0 (0)	1 (1.8)	1.69	0.19
Spirometry
FEV1% predicted†	75.1 [63.95, 85.85]	37.8 [31.75, 44.35]	−10.227	<0.001
FEV1/FVC†	61.67 [55, 66.36]	42.81 [38.2, 49.08]	−8.614	<0.001
MMEF†	31 [24.5, 43]	14.4 [11.05, 16.8]	−8.761	<0.001
GOLD stages			−10.738	<0.001
Stage I: mild (≥80%)	40 (42.6)	0 (0)
Stage II: moderate (50–79%)	54 (57.4)	0 (0)
Stage III: severe (30–49%)	0 (0)	45 (80.4)
Stage IV: very severe (<30%)	0 (0)	11 (19.6)

Data are presented as median [interquartile range], n (%), or mean ± standard deviation. ^†, continuous data without normal distribution. BMI, body mass index; CAD, coronary artery disease; COPD, chronic obstructive pulmonary disease; FEV1% predicted, FEV1 percentage of predicted; FEV1/FVC, forced expiratory volume in 1 second per forced vital capacity; GOLD, Global Initiative for Chronic Obstructive Lung Disease; MMEF, maximal mid-expiratory flow; T2DM, type 2 diabetes.

LASSO regression analysis

LASSO regression was employed to analyze the radiomics features potentially associated with GOLD stages, which revealed eight GOLD stages-associated radiomics features, including (I) original_firstorder_10Percentile; (II) original_shape_MinorAxisLength, original_shape_Elongation; (III) original_shape_elongation; (IV) original_gldm_SmallDependenceLowGrayLevelEmphasis; (V) log_sigma.3.0.mm_3D_firstorder_mean; (VI) wavelet.LLL_firstorder_10Percentile; (VII) wavelet.LHL_glszm_GrayLevelVariance; and (VIII) wavelet.LLH_glcm_Imc1 (Figure 3A-3C).

Figure 3.

Potential variables associated with the GOLD stages of COPD were selected by the LASSO regression model. (A) LASSO coefficient profiles for all variables. (B) Identification of the optimal penalization coefficient (λ) in the LASSO model, which was carried out by 10-fold cross-validation by minimum criteria and 1 − SE criteria. Left line: the minimum error; right line: the cross-validated error within one standard error of the minimum. (C) LASSO coefficient values of eight potential variables. X-axis means the coefficient in the LASSO Model. Y-axis means the eight optimal features selected from 944 features. COPD, chronic obstructive pulmonary disease; GOLD, Global Initiative for Chronic Obstructive Lung Disease; LASSO, least absolute shrinkage and selection operator; SE, standard error.

Binary logistic regression analysis

Subsequently, these eight LASSO regression-selected variables were substituted into the binary logistic regression model. Then, our data identified that two radiomics features, including original_firstorder_10Percentile and wavelet.LHL_glszm_GrayLevelVariance, were independently associated with GOLD stages in the patients with COPD (Table 2). The C-index of this model was 0.838 [95% confidence interval (CI): 0.773–0.903].

Table 2. Binary logistic regression analysis of independent radiomics features associated with GOLD stages in patients with COPD (n=150).

Radiomics features	OR	95% CI	P value
original_firstorder_10Percentile	0.034	0.160–0.754	0.02
wavelet.LLL_firstorder_10Percentile	0.413	1.689–6.909	0.47
wavelet.LHL_glszm_GrayLevelVariance	0.198	0.442–0.988	0.047
wavelet.LLH_glcm_Imc1	0.444	0.747–1.260	0.27
log.sigma.3.0.mm.3D_firstorder_Mean	0.423	0.727–1.249	0.25
original_gldm_SmallDependenceLowGrayLevelEmphasis	0.587	0.980–1.634	0.94
original_shape_MinorAxisLength	0.440	0.928–1.955	0.84
original_shape_Elongation	0.883	1.7098–3.306	0.11

CI, confidence interval; COPD, chronic obstructive pulmonary disease; GOLD, Global Initiative for Chronic Obstructive Lung Disease; OR, odds ratio.

Visualizing and validating the model using a nomogram

Based on the aforementioned binary logistic regression model, a nomogram was constructed to predict clinical outcomes (Figure 4A). The calibration curve, generated with 1000 bootstrap resamples, demonstrated excellent agreement between the predicted and actual probabilities, with both the Apparent and Bias-corrected lines closely approximating the Ideal line [mean absolute error (MAE) =0.037; Figure 4B]. Furthermore, the model exhibited good discriminative ability, achieving an area under the ROC curve (AUC) of 0.820 (95% CI: 0.749–0.891; Figure 4C). The DCA curve demonstrated that when the risk threshold for the nomogram ranged from 0.13 to 0.84, the predicted net benefit of the model surpassed both the ‘All’ and ‘None’ lines (Figure 4D). Moreover, the CIC also showed that the radiomics model showed good predictive performance over the entire range of threshold (Figure 4E).

Figure 4.

The nomogram for predicting GOLD stages in COPD patients. A nomogram was utilized to visualize and verify the binary logistic regression model. (A) Nomogram. The total point of a specific patient is the sum of individual variable points. The predicted probability of GOLD stage is on the GOLD stage-scale, which is corresponding to total points-scale. (B) Calibration curve. Ideal line: the nomogram reference line; apparent line: the actual probability of each patient in our study; bias-corrected line is adjusted by bootstrap with 1,000 resamples. The length of the vertical lines at the top of the plot represents the number of patients. (C) ROC curve. (D) DCA curve. None line: the assumption that all patients had GOLD I–II. All line: the assumption that all patients had GOLD III–IV. Red line: the nomogram. (E) CIC curve. Number high risk line: the number of people who are classified as GOLD III–IV by the model at each threshold probability; number high risk with event line: the number of GOLD III–IV at each threshold probability. CIC, clinical impact curve; COPD, chronic obstructive pulmonary disease; DCA, decision curve analysis; GOLD, Global Initiative for Chronic Obstructive Lung Disease; ROC, receiver operating characteristic.

Discussion

In this cross-sectional study, a total of 223 COPD patients were recruited. Ultimately, 94 GOLD I–II patients and 56 GOLD III–IV patients were included. A set of 944 quantitative radiomics features was derived from the chest CT images for every subject. LASSO regression selected 8 variables, including (I) original_firstorder_10Percentile; (II) original_shape_MinorAxisLength, original_shape_Elongation; (III) original_shape_elongation; (IV) original_gldm_SmallDependenceLowGrayLevelEmphasis; (V) log_sigma.3.0.mm_3D_firstorder_mean; (VI) wavelet.LLL_firstorder_10Percentile; (VII) wavelet.LHL_glszm_GrayLevelVariance; and (VIII) wavelet.LLH_glcm_Imc1. These variables were potentially associated with GOLD stages. Subsequently, these 8 LASSO regression-selected variables were substituted into the binary logistic regression model. We found that 2 radiomics features, including original_firstorder_10Percentile and wavelet.LHL_glszm_GrayLevelVariance, were independently associated with GOLD stages. The validation results collectively confirmed the robustness and clinical utility of our binary logistic regression model. The model attained the AUC of 0.820 (95% CI: 0.749–0.891), indicating good discriminative ability. Calibration was excellent, with an MAE of 0.037. DCA further demonstrated the model’s clinical net benefit across a wide threshold probability range (0.13 to 0.84), a finding corroborated by the CIC. Therefore, the nomogram and its associated validation plots verified the model’s accuracy, reliability, and potential value as a promising tool for predicting GOLD stages in COPD patients.

COPD, a heterogeneous chronic lung disease, is characterized by a range of abnormalities, such as small airway remodeling, pulmonary vascular remodeling, and emphysema (2,23). Although PFT remains the clinical gold standard for diagnosis, CT plays an indispensable role in COPD management due to its high precision and visual evaluation capabilities (24,25). Although the incidence and disease burden of COPD in China is heavy, the lung function testing rate is still insufficient in COPD management (3). Simultaneously, the utilization of chest CT is notably high, particularly due to the widespread implementation of large-scale chest CT examinations and pulmonary nodule screenings. Meanwhile, an increasing number of community health service centers are being equipped with CT technology. Consequently, chest CT enhances the detection of COPD, which can help to reduce the social and economic burden as well as improve the quality of life. Recent study demonstrated that CT-based imaging features, derived solely from inspiratory CT scans, outperformed existing advanced methods in detecting COPD, both in standard and low-dose CT scans (26).

Radiomics holds significant potential for extracting valuable medical information and improving the accuracy of clinical differential diagnosis (27,28). Several studies revealed the values of lung imaging feature, derived from CT scans and clinical manifestations, in the diagnosis and severity assessment of COPD (22,29,30). As a quantitative methodology, radiomics offered a systematic approach to analyzing medical images. Therefore, the primary objective of this study was to evaluate the role of radiomics in assessing COPD severity.

We extracted 944 potential radiomic features from the CT images in this study. These features encompassed first-order statistics, which were derived through histogram analysis, which included commonly used and fundamental metrics, encompassing energy, entropy, mean, and median. Additionally, textural features were computed from various matrices, including the Gray Level Dependence Matrix (GLDM), Gray Level Size Zone Matrix (GLSZM), Grey Level Co-occurrence Matrix (GLCM), and Grey Level Run-Length Matrix (GLRLM). High-order filter transform features were further acquired. These features included both intensity-based and texture-based characteristics, derived via the application of filter transformations and wavelet transformations to the original images. These transformations were performed using a range of filters, including wavelet-LLL, wavelet-HLL, wavelet-LHL, wavelet-LLH, logarithmic, square root, square, original, and exponential. First-order features represented the most basic information contained within an image. These features were derived from the grayscale histogram of the image, which provided the distribution and frequency of pixels with specific intensities in the region of interest. Texture features, encompassing gray-level co-occurrence matrices and gray-level correlation matrices, delineated the spatial relationships between each pixel and its neighboring pixels. Wavelet features primarily captured the time-frequency domain characteristics within the lesion (31). Collectively, these features reflected information from pixel intensity and texture morphology that could not be discerned by the naked eye of a radiologist (32). In this study, 944 imaging features were initially selected and subsequently screened using LASSO, resulting in the identification of eight GOLD stages potentially associated features. These features were incorporated into a binary logistic regression model. Then, two features were found to be independently associated with the GOLD stages of COPD, which were original_firstorder_10Percentile and wavelet.LHL_glszm_graylevelvariance. It is imperative to highlight that the radiologist is unable to discern these imaging characteristics with the eyes and can only be identified using radiomics analysis (33). These results indicate that radiomics can extract additional disease-related information and assist in clinical diagnosis via CT imaging.

This study was a prospective cross-sectional study utilizing high-quality patient data. However, several limitations should be acknowledged. Firstly, in our study, the radiomics methodology employed was not most advanced. Specifically, features were sampled from 11 levels of both lungs rather than being extracted from the entire lung. Secondly, the sample size was relatively limited. Thirdly, the study was not conducted across multiple centers. Therefore, future research will aim to refine feature extraction techniques, increase the sample size, and incorporate multi-center collaborations.

Conclusions

Collectively, we identified a noticeable difference in radiomics features between GOLD I–II and GOLD III–IV patients with COPD. The result indicates that the CT-based radiomics features, original_firstorder_10Percentile and wavelet.LHL_glszm_GrayLevelVariance, can be potentially used to predict the severity of COPD. However, our results require further validation through multicenter and large-scale clinical studies. Radiomics serves as a robust indicator for assessing the severity of COPD and holds significant potential in guiding clinical management. Further research is warranted to enhance the classification accuracy of COPD severity.

Supplementary

The article’s supplementary files as

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 Research Ethics Committees of the Suining Central Hospital (No. KYLLKS20230103) and individual consent for this retrospective analysis was waived.

Data Sharing Statement

Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1-2721/dss

jtd-18-03-235-dss.pdf ^{(99.3KB, pdf)}

DOI: 10.21037/jtd-2025-1-2721