The value of CT-based radiomics for differentiation of pleural effusions in bacterial pneumonia and Mycoplasma pneumoniae pneumonia in children

Introduction

Pleural effusion (PE) is a common complication that can be caused by a variety of etiologies. In children, PE is usually caused by infectious factors, especially community-acquired pneumonia (1). Pneumonia is one of the leading causes of death in children under five years old (2), and it has been shown that children with PE have a significantly increased risk of death compared to those with a normal chest X-ray (3). PE not only affects pulmonary ventilation function and hemodynamic stability (4), but may also worsen symptoms, prolong the disease course and increase the length of hospitalization in children with pneumonia (5). Therefore, early diagnosis and treatment of PE are essential to improve the prognosis of pediatric patients.

Bacterial pathogens (such as Streptococcus pneumoniae) and Mycoplasma pneumoniae are the two main pathogens that cause community-acquired pneumonia complicated with PE in children (5-7). Although pathogen culture using blood, respiratory specimens or PE aspirated by thoracentesis can be used for definitive diagnosis, these methods have disadvantages such as low positivity rates and time-consuming procedures (8). Furthermore, thoracentesis is invasive and not applicable to all children, especially those with only small amounts of PE. In contrast, computed tomography (CT) examination can more sensitively and accurately detect changes in the pleura and pleural cavity, but accurately differentiating between bacterial pneumonia PE (BPPE) and Mycoplasma pneumoniae parapneumonic effusion (MPPE) remains challenging. In view of these considerations, finding a rapid, effective and non-invasive identification method is of great significance in guiding antibiotic therapy, reducing the emergence of drug-resistant microorganisms, improving the efficacy of treatment, shortening the disease course and alleviating the suffering in children.

Radiomics provides a new approach to the diagnosis of diseases by deeply analyzing medical images to extract a large amount of high-throughput information that often cannot be identified and quantified by radiologists through the naked eye. Currently, no studies have reported the use of CT-based radiomics to differentiate between BPPE and MPPE in children. Therefore, the aim of this study was to explore the feasibility and value of a radiomics approach based on non-contrast chest CT scans in the differentiation of BPPE and MPPE in children, and to construct a machine learning model to effectively differentiate between BPPE and MPPE, thereby providing crucial support for enhancing the accuracy of imaging diagnosis, facilitating early clinical diagnosis and treatment, and improving the prognosis of pediatric patients. This manuscript is written accordance to the TRIPOD reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-24-364/rc).

Methods

Clinical information

The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This study was approved by the Ethics Committee of the Children’s Hospital of Chongqing Medical University (No. 2024-289), and informed consent was waived for this retrospective study. We retrospectively collected clinical and CT imaging data of hospitalized children with PE detected by chest CT scans from December 2020 to December 2023 in the Children’s Hospital of Chongqing Medical University.

The inclusion criteria were as follows: (I) bacterial infection: a positive sputum, blood, pleural fluid or bronchoalveolar lavage fluid (BALF) culture, or a positive nucleic acid test for lower respiratory tract bacteria, and effective treatment with antibiotics; (II) Mycoplasma pneumoniae infection: a positive polymerase chain reaction (PCR) in sputum, BALF or pleural fluid, and/or serologic test for Mycoplasma pneumoniae (semi-quantitative) ≥1:160.

Exclusion criteria: (I) bacterial and Mycoplasma pneumoniae coinfection; (II) tuberculosis, parasitic and other pathogenic infections; (III) non-infectious PE; (IV) poor CT image quality; (V) minimal amounts of PE making it difficult to segment images. The screening process is shown in Figure 1.

Figure 1 Patient screening workflow for this study. PE, pleural effusion; CT, computed tomography; BALF, bronchoalveolar lavage fluid; PCR, polymerase chain reaction.

Patient grouping: a total of 535 patients were included, including 167 patients with BPPE and 368 patients with MPPE. The BPPE group consisted of 98 males and 69 females, with a mean age of 3.6 years (range, 14 days–16 years). The MPPE group consisted of 187 males and 181 females, with a mean age of 7.2 years (range, 4 months–15 years). All patients were randomly divided into a training set and a test set in the ratio of 7:3. The training set comprised 117 cases of BPPE and 257 cases of MPPE. The test set included 50 cases of BPPE and 111 cases of MPPE.

CT examinations

Children underwent chest CT scans in a quiet state. For those who were unable to cooperate, appropriate sedation measures were taken to ensure a smooth examination. A GE LightSpeed VCT (General Electric Company, Boston, USA) 64-row spiral CT or a Philips Brilliance iCT (Philips Healthcare, Amsterdam, the Netherlands) 256-row spiral CT was used for scanning, with the following scanning parameters: tube voltage of 75–110 kV, automated tube current modulation, layer thickness of 5.0 mm, layer spacing of 5 mm, pitch of 0.984:1, and a matrix of 512×512. The scanning range was from the apex to the base of the lung.

Data pre-processing and image segmentation

The images of the patients’ chest CT scans without contrast were exported from the Picture Archiving and Communication System (PACS) in Digital Imaging and Communications in Medicine (DICOM) format, and processed in the uAI Research Portal software (version 202311115, Shanghai United Imaging Intelligence Medical Technology Co., Ltd., Shanghai, China), which is compliant with the Image Biomarker Standardization Initiative. In order to reduce the influence of different machines and scanning parameters, CT images were first pre-processed by normalization (window level of 50 HU, window width of 350 HU) and image re-sampling. Then a radiologist with 5 years of experience sketched the region of interest (ROI) layer by layer along the edge of PE to generate volume of interest (VOI). And a radiologist with more than 15 years of experience reviewed and verified the accuracy of all ROIs. All images were outlined on the mediastinal window, trying to avoid outlining adjacent lung lesions, intercostal muscles, and other normal anatomical structures (Figure 2). To ensure consistency of delineation, 50 cases were randomly selected for re-outlining after one month by the junior radiologist.

Figure 2 Delineation of ROI for a patient (7 years old, male) with Mycoplasma pneumoniae pneumonia pleural effusion: (A) original CT image; (B) ROI on one slice; (C) volume of interest formed after delineating each slice. Delineation of ROI for a patient (4 years old, male) with bacterial pneumonia pleural effusion: (D) original CT image; (E) ROI on one slice; (F) volume of interest formed after delineating each slice. ROI, region of interest; CT, computed tomography.

Feature extraction

All radiomics features were extracted in the uAI Research Portal software. Seven types of radiomics features were extracted from the original images and filtered images, including first-order features, shape-based features, gray level co-occurrence matrix (GLCM) features, gray level run length matrix (GLRLM) features, gray level size zone matrix (GLSZM) features, gray level dependence matrix (GLDM) features, neighboring gray tone difference matrix (NGTDM) features. The filters used in this study included box mean, additive Gaussian noise, binomial blur image, curvature flow, box sigma image, Laplacian of Gaussian (LoG), wavelet, normalize, Laplacian sharpening, discrete Gaussian, mean, speckle noise, recursive Gaussian and shot noise. The intra-class correlation coefficient (ICC) was calculated. Radiomics features with ICC >0.75 were selected to ensure consistency and reliability of the features.

Feature selection

A large number of extracted radiomics features must be screened to retain the most relevant features for the identification of BPPE and MPPE. All features with ICC >0.75 were normalized by z-score. The Select K Best (K value =20) was used first. Subsequently, the max-relevance and min-redundancy (mRMR) was utilized to reduce redundancy and select the features with the highest relevance for the classification of PE, with the features retained by 10. Finally, the least absolute shrinkage and selection operator (LASSO) was utilized to select the optimal radiomics features.

Model construction and evaluation

Based on the final selection of optimal radiomics features, logistic regression (LR) was chosen to construct the radiomics model. To evaluate the model performance, the receiver operating characteristic (ROC) curves were plotted, and key metrics such as area under the curve (AUC), 95% confidence interval (CI), sensitivity, specificity, and accuracy were calculated. The calibration curves were plotted to more intuitively demonstrate how well the actual predictive ability of the model matched the theoretical predictive ability. The decision curve analysis (DCA) was utilized to evaluate the clinical effectiveness of the model in PE classification by quantifying the net benefits under different threshold probabilities.

Statistical analysis

The model performance was evaluated using ROC curves in both the training and test sets, using evaluation metrics such as AUC, sensitivity and specificity.

Results

Radiomics features

A total of 2,264 radiomics features were extracted from ROI of each patient, and 2,046 radiomics features with ICC >0.75 were selected, including 18 first-order features, 14 shape-based features, 20 GLCM features, 16 GLRLM features, 15 GLSZM features, 12 GLDM features, 5 NGTDM features, and 1,946 higher-order statistical features obtained by filtering and transforming the original image. Seven optimal features were finally selected by Select K Best, mRMR, and LASSO, mainly consisting of first-order features, GLSZM features, and GLDM feature (Figure 3A). To analyze the interrelationships among these features, Figure 3B shows the correlation coefficients among the optimal radiomics features in the training set.

Figure 3 The final selected radiomics features and their coefficients and correlations. (A) The coefficients for the final selected radiomics features. (B) The correlation map of the final selected radiomics features. The numbers on the right side illustrate the colors corresponding to the different correlation coefficients.

Radiomics model performance

The ROC curves of the model are shown in Figure 4, the AUC in the training set was 0.942 (95% CI: 0.917–0.967), with sensitivity, specificity, accuracy and precision of 89.9%, 82.1%, 87.4% and 91.7%, respectively. The AUC in the test set was 0.917 (95% CI: 0.868–0.965), with sensitivity, specificity, accuracy and precision of 87.4%, 80.0%, 85.1% and 90.7%, respectively (Table 1). These results indicated that the LR model had good performance in differentiating between BPPE and MPPE.

Figure 4 Receiver operating characteristic curves of the model. ROC, receiver operating characteristic; AUC, area under the curve.

The calibration curves for the training and test sets are shown in Figure 5. The DCA (Figure 6) showed excellent net benefits at different thresholds, suggesting a good clinical utility of the model.

Figure 5 Calibration curves of the model.

Figure 6 Decision curves of the model.

Discussion

In this study, we explored the value of CT radiomics in differentiating BPPE and MPPE in children. Currently, although all imaging examinations cannot accurately differentiate between BPPE and MPPE, radiomics has demonstrated great potential in disease diagnosis, classification, and prognosis prediction (9-13). The results of our study suggest that the features extracted by the radiomics method can effectively differentiate between BPPE and MPPE, which is consistent with the findings of Cai et al. (14) in the differentiation of benign and malignant PE in adults. In addition, our study also found that the radiomics model alone demonstrated good discriminatory ability even in the absence of clinical features, emphasizing the importance of radiomics features in diagnosis.

In this study, we retrospectively analyzed the data of hospitalized children with PE detected by CT scans in the past three years. We found that there were fewer children with BPPE than with MPPE, with the following possible reasons: (I) as a tertiary medical institution, many children had already received antibiotics prior to admission, which may have affected the detection rate of bacterial pathogens (15,16); (II) the incidence of Mycoplasma pneumoniae pneumonia has shown an upward trend after the pandemic of coronavirus disease 2019 (COVID-19), which may have contributed to an increase in the number of children with MPPE (17,18). Furthermore, a study in the Northeast and Inner Mongolia of China (19) also showed a dramatic increase in Mycoplasma pneumoniae infections among children, with a significant increase in cases of complicated by PE.

In our study, we ultimately identified seven optimal radiomics features through Select K Best, mRMR and LASSO, including first-order features and texture features, which can effectively differentiate between BPPE and MPPE. These findings are in agreement with those of Han et al. (20), whose study confirmed the effectiveness of first-order features and texture features (including GLSZM, GLDM and GLCM) in differentiating malignant from benign PE. Our study further demonstrates that these features are able to reflect differences in gray level that cannot be recognized by the naked eye in PE, thus quantifying the heterogeneity of the lesions. First-order features mainly reflect the distribution of voxel intensities within the ROI, without reflecting the relative positional relationships (21). First-order features have good reproducibility (22,23) and can reflect the overall density of the tumor (24). The maximum in first-order features can quantify the heterogeneity within the tumor (25). In a study applying CT radiomics features to construct a nomogram to differentiate primary pulmonary tuberculosis from community-acquired pneumonia in children, first-order features (mainly the maximum) extracted from pulmonary consolidation were found to be useful in differentiating the two lesions (26). A similar finding was made in our study, namely that the maximum was one of the important features to discriminate between BPPE and MPPE. The maximum is the highest gray level intensity within the ROI, we speculate that the undetectable difference of gray level in PE by the naked eye may reflect the variation of PE density, allowing to quantify the heterogeneity of lesions and thereby discriminate between BPPE and MPPE. Texture features mainly provide information about the spatial location of voxels distribution in images (21). They can quantify the gray level, coarseness and homogeneity of lesions that are indistinguishable by the naked eye (27). The texture features used to construct the model in our study were GLSZM and GLDM. The GLSZM features quantify gray level zones in the ROI, which are the number of connected voxels with the same gray level intensity. They describe the homogeneity and heterogeneity of the ROI (28), indicating that there are differences in texture homogeneity between BPPE and MPPE. The GLDM features quantify gray level dependencies in the ROI, which are the number of connected voxels that depend on the center voxel within a certain distance (28), they can reflect the gray level inhomogeneity in the whole lesion (29). Meanwhile, the optimal features in our study were mainly derived from filter transformations. We found that there were duplicate features, but they were obtained from different filters. Since various filters have different principles and mechanisms of action, the final features obtained were different. The features transformed by filters can help to reveal information that is not visible to the naked eye and is valuable for analyzing the lesion (30,31).

The LR model used in this study is a classical machine learning model that predicts classification results by calculating probability values and provides a better explanation of the relationship between features and target variables (32). The LR model achieved excellent discriminatory performance in this study, with an AUC value of 0.917 in the test set and with high sensitivity, specificity and accuracy.

Despite this study achieved positive results, there are some limitations. Firstly, as a retrospective study, the sample size may have selection bias, and prospective studies are needed to validate the reliability of the model. Secondly, this study was conducted at a single center and lacked external validation, it is necessary to expand the sample size and incorporate data from multiple centers to enhance and validate the reliability of the model in the future. Additionally, this study only used LR to build the classification model, future studies should explore the efficacy of other machine learning classification methods and consider incorporating clinical data to further improve the model.

Conclusions

In conclusion, this study demonstrates the potential of CT radiomics in differentiating BPPE and MPPE in children and provides a new direction for future research. Nevertheless, there are few studies on PE using radiomics, and further research and validation are necessary to facilitate its clinical application. Additionally, the relevant biological interpretations of radiomic features require further study and exploration. We believe that with continuous advancement of technology and in-depth research, radiomics will offer greater support for clinical diagnosis and treatment.

Acknowledgments

Funding: This study was supported by the Joint Fund for Innovation and Development of the Natural Science Foundation of Chongqing, China [No. (CSTB)2024NSCQ-LZX0090].

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tp.amegroups.com/article/view/10.21037/tp-24-364/rc

Data Sharing Statement: Available at https://tp.amegroups.com/article/view/10.21037/tp-24-364/dss

Peer Review File: Available at https://tp.amegroups.com/article/view/10.21037/tp-24-364/prf

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tp.amegroups.com/article/view/10.21037/tp-24-364/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 (as revised in 2013). The study was approved by the Ethics Committee of the Children’s Hospital of Chongqing Medical University (No. 2024-289) and informed consent was waived for this retrospective analysis.

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. Krenke K, Sadowy E, Podsiadły E, et al. Etiology of parapneumonic effusion and pleural empyema in children. The role of conventional and molecular microbiological tests. Respir Med 2016;116:28-33. [Crossref] [PubMed]
2. World Health Organization [Internet]. Pneumonia in Children; 2024 [cited 2024 Sept 5]. Available online: https://www.who.int/news-room/fact-sheets/detail/pneumonia
3. Wilkes C, Bava M, Graham HR, et al. What are the risk factors for death among children with pneumonia in low- and middle-income countries? A systematic review. J Glob Health 2023;13:05003. [Crossref] [PubMed]
4. Brogi E, Gargani L, Bignami E, et al. Thoracic ultrasound for pleural effusion in the intensive care unit: a narrative review from diagnosis to treatment. Crit Care 2017;21:325. [Crossref] [PubMed]
5. de Benedictis FM, Kerem E, Chang AB, et al. Complicated pneumonia in children. Lancet 2020;396:786-98. [Crossref] [PubMed]
6. Lohuis SJ, de Groot E, Kamps AWA, et al. Conservative Treatment of Parapneumonic Effusion in Children: Experience From a 10-Year Consecutive Case Series. Pediatr Infect Dis J 2023;42:180-3. [Crossref] [PubMed]
7. Kutty PK, Jain S, Taylor TH, et al. Mycoplasma pneumoniae Among Children Hospitalized With Community-acquired Pneumonia. Clin Infect Dis 2019;68:5-12. [Crossref] [PubMed]
8. Bradley JS, Byington CL, Shah SS, et al. The management of community-acquired pneumonia in infants and children older than 3 months of age: clinical practice guidelines by the Pediatric Infectious Diseases Society and the Infectious Diseases Society of America. Clin Infect Dis 2011;53:e25-76. [Crossref] [PubMed]
9. Chen J, Yang F, Liu C, et al. Diagnostic value of a CT-based radiomics nomogram for discrimination of benign and early stage malignant ovarian tumors. Eur J Med Res 2023;28:609. [Crossref] [PubMed]
10. Goda JS, Punwatkar D, Jha A, et al. CT Based Tumor Radiomics with Machine Learning Classifiers for Molecular Subtyping of Diffuse Large B Cell Lymphoma. Int J Radiat Oncol Biol Phys 2023;117:e467.
11. Ye JY, Fang P, Peng ZP, et al. A radiomics-based interpretable model to predict the pathological grade of pancreatic neuroendocrine tumors. Eur Radiol 2024;34:1994-2005. [Crossref] [PubMed]
12. Wang H, Li T, Xie M, et al. Association of Computed Tomography Radiomics Signature with Progression-free Survival in Neuroblastoma Patients. Clin Oncol (R Coll Radiol) 2023;35:e639-47. [Crossref] [PubMed]
13. Wang H, Chen X, Li T, et al. Identification of an Ultra-High-Risk Subgroup of Neuroblastoma Patients within the High-Risk Cohort Using a Computed Tomography-Based Radiomics Approach. Acad Radiol 2024;31:1655-65. [Crossref] [PubMed]
14. Cai F, Cheng L, Liao X, et al. An Integrated Clinical and Computerized Tomography-Based Radiomic Feature Model to Separate Benign from Malignant Pleural Effusion. Respiration 2024;103:406-16. [Crossref] [PubMed]
15. Stankey CT, Spaulding AB, Doucette A, et al. Blood Culture and Pleural Fluid Culture Yields in Pediatric Empyema Patients: A Retrospective Review, 1996-2016. Pediatr Infect Dis J 2018;37:952-4. [Crossref] [PubMed]
16. Forster J, Piazza G, Goettler D, et al. Effect of Prehospital Antibiotic Therapy on Clinical Outcome and Pathogen Detection in Children With Parapneumonic Pleural Effusion/Pleural Empyema. Pediatr Infect Dis J 2021;40:544-9. [Crossref] [PubMed]
17. Qiu W, Ding J, Zhang H, et al. Mycoplasma pneumoniae detections in children with lower respiratory infection before and during the COVID-19 pandemic: a large sample study in China from 2019 to 2022. BMC Infect Dis 2024;24:549. [Crossref] [PubMed]
18. Yan C, Xue GH, Zhao HQ, et al. Current status of Mycoplasma pneumoniae infection in China. World J Pediatr 2024;20:1-4. [Crossref] [PubMed]
19. Wang F, Cheng Q, Duo H, et al. Childhood Mycoplasma pneumoniae: epidemiology and manifestation in Northeast and Inner Mongolia, China. Microbiol Spectr 2024;12:e0009724. [Crossref] [PubMed]
20. Han R, Huang L, Zhou S, et al. Novel clinical radiomic nomogram method for differentiating malignant from non-malignant pleural effusions. Heliyon 2023;9:e18056. [Crossref] [PubMed]
21. Rogers W, Thulasi Seetha S, Refaee TAG, et al. Radiomics: from qualitative to quantitative imaging. Br J Radiol 2020;93:20190948. [Crossref] [PubMed]
22. Desseroit MC, Tixier F, Weber WA, et al. Reliability of PET/CT Shape and Heterogeneity Features in Functional and Morphologic Components of Non-Small Cell Lung Cancer Tumors: A Repeatability Analysis in a Prospective Multicenter Cohort. J Nucl Med 2017;58:406-11. [Crossref] [PubMed]
23. Huynh E, Coroller TP, Narayan V, et al. Associations of Radiomic Data Extracted from Static and Respiratory-Gated CT Scans with Disease Recurrence in Lung Cancer Patients Treated with SBRT. PLoS One 2017;12:e0169172. [Crossref] [PubMed]
24. Chen Z, Yi L, Peng Z, et al. Development and validation of a radiomic nomogram based on pretherapy dual-energy CT for distinguishing adenocarcinoma from squamous cell carcinoma of the lung. Front Oncol 2022;12:949111. [Crossref] [PubMed]
25. Aerts HJ, Velazquez ER, Leijenaar RT, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun 2014;5:4006. [Crossref] [PubMed]
26. Wang B, Li M, Ma H, et al. Computed tomography-based predictive nomogram for differentiating primary progressive pulmonary tuberculosis from community-acquired pneumonia in children. BMC Med Imaging 2019;19:63. [Crossref] [PubMed]
27. Incoronato M, Aiello M, Infante T, et al. Radiogenomic Analysis of Oncological Data: A Technical Survey. Int J Mol Sci 2017;18:805. [Crossref] [PubMed]
28. Scapicchio C, Gabelloni M, Barucci A, et al. A deep look into radiomics. Radiol Med 2021;126:1296-311. [Crossref] [PubMed]
29. Mayerhoefer ME, Materka A, Langs G, et al. Introduction to Radiomics. J Nucl Med 2020;61:488-95. [Crossref] [PubMed]
30. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016;278:563-77. [Crossref] [PubMed]
31. Zhou T, Tu W, Dong P, et al. CT-Based Radiomic Nomogram for the Prediction of Chronic Obstructive Pulmonary Disease in Patients with Lung cancer. Acad Radiol 2023;30:2894-903. [Crossref] [PubMed]
32. Zhao J, Zhan Y, Zhou Y, et al. CT-based radiomics research for discriminating the risk stratification of pheochromocytoma using different machine learning models: a multi-center study. Abdom Radiol (NY) 2024;49:1569-83. [Crossref] [PubMed]