Introduction

Esophageal cancer is a common malignant tumor of the digestive tract, and radiotherapy is one of the important treatment measures.1,2 In clinical practice, the dosage of radiotherapy is an important factor determining the efficacy of non-surgical treatment for esophageal cancer patients.3 However, as the radiation dose increases, the adverse reactions of esophageal cancer patients also increase. During the process of receiving radiation therapy for patients, the high sensitivity of normal lung tissue to radiation limits the application of radiation therapy in chest tumors to a certain extent.4 Previous studies have shown that the incidence of radiation-induced lung injury in chest malignant tumors after radiotherapy ranges from 10% to 30%, and the actual number of patients who experience radiation-induced lung injury in clinical practice is underestimated.5–7

In clinical practice, radiation-induced lung injury can be divided into acute radiation-induced pneumonitis and long-term radiation-induced pulmonary fibrosis.7 Among them, radiation pneumonitis often occurs within 3 months after patients receive radiation therapy, with clinical manifestations including cough, difficulty breathing, and in severe cases, it can further develop into respiratory failure, even endangering the patient’s life. At present, the pathogenesis of radiation-induced lung disease is not yet clear, and there is a lack of targeted treatment measures.8 In view of this, relying on certain early indicators or non-invasive examinations to predict the incidence of radiation pneumonitis has important clinical guidance value for designing personalized treatment plans for high-risk patients.

At present, the parameters used to predict radiation pneumonitis in clinical practice include general clinical information of patients, pulmonary dosimetry parameters, and changes in lung function, etc.9–11 These indicators have their own advantages and disadvantages in predicting radiation pneumonitis. In recent years, with the rapid development of radiomics and its widespread use in auxiliary diagnosis and prognosis prediction, it has provided the possibility for screening reference predictive factors and improving the predictive efficiency of predicting radiation pneumonitis.12 In addition, with the continuous application of deep learning in clinical medicine, the diagnosis and treatment of a large number of diseases rely on advanced machine learning algorithms to better serve clinical decision-making.13 Encouraged by this, this study intends to use radiomics methods to capture features from CT images of esophageal cancer patients, and construct a machine learning prediction model by integrating the dosimetric parameters of the patient’s affected lung, in order to explore the predictive value of radiation pneumonitis and guide clinical practice more reasonably.

Materials and MethodsStudy Population

We retrospectively included 228 esophageal cancer patients who were admitted to the radiotherapy department of our hospital from January 2017 to December 2023 and were unable to undergo surgical treatment but received radiotherapy for the first time. Inclusion criteria: 1) Patients diagnosed with esophageal squamous cell carcinoma by pathology; 2) Patients with Karnofsky score ≥70; 3) Patients with no abnormal liver or kidney function and no history of chest radiation therapy; 4) Patients with complete dose volume parameters and targeted treatment area located in the chest cavity; 5) Patients who meet the indications for intensity modulated radiation therapy in the trial. Exclusion criteria: 1) Patients with underlying lung diseases such as chronic obstructive pulmonary disease and interstitial lung disease; 2) Patients with low bone marrow hematopoietic function; 3) Patients who have been interrupted by radiotherapy for more than 7 days or do not have a complete radiotherapy plan. This study has been approved by the Medical Ethics Committee of Wuhan First Hospital (NO.202031) and study complies with the Declaration of Helsinki. The personal privacy of all patients included in this study has been kept confidential, and the informed exemption consent of patients for this study has been approved by the ethics committee for implementation. The patient inclusion and prediction model construction process for this study was shown in Figure 1.

Figure 1 Flow chart of patient collection and data processing in this study.

CT Image Acquisition and Quality Control Evaluation

All patients received Siemens performing localization scanning using Go.Now type CT. The scanning voltage is 110kV, the tube current is in automatic mode, the layer thickness is set to 5mm, and the matrix size is 512 × 512 pixels. After the simulation positioning is completed, two physicians with over 10 years of radiation therapy experience can draw the target area and organs at risk in the Monaco planning system. All patient radiotherapy plans are designed by the same radiotherapy physicist, using fixed field intensity modulated radiotherapy technology with 6MV X-ray energy. The radiation field includes two pairs of tangent fields and a zero degree field. The radiotherapy prescription is 50Gy/25fx. All plans are implemented after being reviewed by radiologists and physicists. During radiotherapy, regularly record the values of V5, V10, V15, V20, V25, V30, and mean lung dose (MLD) in the affected lung.

Image Segmentation and Feature Extraction

We imported the pre radiotherapy positioning CT images of patients into 3D Slicer (version 4.11) software in DICOM format, and used a semi-automatic and manual layer-by-layer correction method to segment the affected lung as the region of interest. The segmentation of the patient’s affected lung was completed by an imaging specialist and reviewed by a diagnostic physician with more than 10 years of experience. We used the radiomics plugin in 3D-Slicer software to extract imaging omics features, including original image features, texture features, and intensity and texture features calculated through wavelet decomposition. Before feature extraction, all images were uniformly adjusted to 1 × 1 × 1 pixels using resampling techniques.

Feature Selection and Prediction Model Construction

This study conducted stratified random sampling on patients and divided them into training and testing sets in a 7:3 ratio. Firstly, we calculated the Spearman correlation coefficients for all features and remove feature factors with absolute correlation coefficients greater than 0.9. Then, we used the method of feature recursive elimination to filter and obtain the best set of features. This study used random forests, generalized linear models, decision trees, and neural network algorithms to construct machine learning models. In addition, to avoid overfitting, we used grid search in the training set to facilitate the selection of optimal model parameters, and used 10-fold cross validation for model robustness training and validation.

Statistical Analysis

We used SPSS 26.0 software to conduct a global test on the dosimetric parameters of the affected lung. The t-test was used for inter-group comparisons that followed a normal distribution, while the Wilcoxon rank sum test was used for those that did not. In addition, we used R software (version 4.0.2) for feature screening and machine learning model creation, and evaluated the predictive performance of the model through receiver operator characteristic curve (ROC), as well as evaluated the clinical benefits of different predictive models using decision curve analysis (DCA). P < 0.05 indicates a statistically significant difference.

ResultsPatient Clinical and Treatment Characteristics

A total of 228 patients with esophageal cancer who received radiotherapy were included, including 54 patients diagnosed with radiation pneumonitis (ie grade ≥ 2). These patients were randomly divided into a training group (n = 160) and a testing group (n = 68). In addition, patient clinical data and total dose, 90% planned target area received dose, corresponding proportion of 50Gy irradiation plan target area, bilateral lung volume, lung dose volume parameters V5~V40, and lung average dose waiting for selection parameters were compared between groups. It was found that clinical staging, radiotherapy mode, and radiotherapy dose were correlated with the occurrence of radiation pneumonitis in esophageal cancer patients after radiotherapy (P values < 0.05). The general information and dose-volume parameters of all patients were summarized in Table 1 and Supplementary Table 1.

Table 1 The Characteristics of Study Population and Clinical Outcome

Feature Selection

A total of 23 features were obtained from radiomics and dose-volume parameter analysis. We used the LASSO regression model to select eight features, and all radiomics features used wavelet filters, as shown in Figure 2A and B. These features include esophageal PTVD90, esophageal PTVV50, total dose, MLD, V5, V10, V15, V20. The weights selected by LASSO regression analysis for the radiomics and dosimetry parameter characteristics related to radiation pneumonitis are shown in Supplementary Table 2.

Figure 2 Radiomics features and dose-volume parameter associated with radiation pneumonitis. (A) Spearman correlation analysis; (B) LASSO regression analysis.

Radiation Pneumonitis Nomogram Construction

According to univariate analysis, as shown in Table 2, the clinical staging, esophageal PTVD90, esophageal PTVV50, total dose, MLD, V5, V10, V15, V20 were potential high-risk factors for radiation pneumonitis. Multivariate analysis showed that the esophageal PTVD90, total dose, MLD, V5, V10, V15, V20 were independent predictors of radiation pneumonitis. On the basis of multivariate analysis, we developed a visible radiomics nomogram (Figure 3A) by combining esophageal PTVD90, MLD, V5, V10, V15, V20. In view of this, we also used calibration curves to confirm the consistency between the predicted and actual observed values of the nomogram for radiation pneumonitis. The Hosmer Lemeshow test showed no statistical difference between the predicted and actual values (P = 0.921) (Figure 3B).

Table 2 Univariate and Multivariate Logistic Regression Analysis of Risk Factors for Radiation Pneumonitis

Figure 3 Nomogram visualization model for predicting radiation pneumonitis. (A) Nomogram; (B) Calibration curve.

Radiation Pneumonitis Machine Learning Based Prediction Model

As for the random forest prediction model, when we set mtry = 5 and ntree = 1200, the model will have the lowest error, that is, selecting esophageal PTVD90, esophageal PTVV50, MLD, V5, V10, V15, V20 as explanatory variables. The average decrease accuracy and average decrease Gini coefficient of each variable are shown in Figure 4A and Supplementary Table 3. Based on the average accuracy reduction index, we construct a decision tree using the top 10 variables. In this decision tree, postoperative menstruation; V5, V10, V15, and V20 was classified as a nodal variable for the occurrence of radiation pneumonitis after radiotherapy (Figure 4B). In addition, using the multi-layer perceptron module of the neural network, the input neurons of the model were sequentially inputted, and the occurrence of radiation pneumonitis was taken as the output neuron, as shown in Figure 4C. Finally, based on the artificial neural network prediction model, the normalized importance ranking of each dependent variable included esophageal PTVD90, esophageal PTVV50, MLD, V5, V10, V15, V20.

Figure 4 Machine learning based visualization model for predicting radiation pneumonitis. (A) Random forest model; (B) Decision tree model; (C) Artificial neural network model.

Performance Metrics of Four Machine Learning Prediction Models

As shown in Table 3, in the validation and testing queue, the random forest prediction model showed AUC values of 0.891 (95% CI:0.840~0.942) and 0.887 (95% CI: 0.836~0.938), respectively. In the validation and testing queue, the AUC of the neural network prediction model was 0.882 (95% CI: 0.831~0.933) and 0.879 (95% CI: 0.828~0.930), respectively. In contrast, the decision tree prediction model had AUC values of 0.811 (95% CI: 0.760~0.862) and 0.809 (95% CI: 0.758~0.860) in the training and validation sets, respectively. Obviously, as shown in Figure 5, the DCA showed that decision tree predictive ability was worse than that of random forests and neural networks, but it was better than the generalized linear prediction model (ie, the Nomogram) with AUC values of 0.751 (95% CI: 0.700~0.802) and 0.762 (95% CI: 0.711~0.813) in the training and validation sets, respectively.

Table 3 Comparison of Predictive Performance of Four Types of Radiation-Induced Pneumonitis Prediction Models Through ROC Analysis

Figure 5 Comparing the predictive performance of machine learning prediction models using DCA. (A) Training set; (B) Validation set.

The Interpretability of the Optimal Radiation Pneumonitis Prediction Model

Due to the best predictive performance of the random forest prediction model, in order to test its predictive performance in clinical practice. We conducted simulation training using clinical impact curve (CIC), as shown in Figure 6. In terms of distinguishing the risk of radiation pneumonitis from non-radiation pneumonitis, the two can achieve the maximum span of differentiation. In view of this, using machine learning techniques from random forests can establish more effective prediction models. This predictive model may be useful in clinical practice and help clinicians tailor precise management and treatment for radiation pneumonitis in esophageal cancer patients.

Figure 6 Differentiation performance of random forest prediction models based on CIC evaluation. (A) Training set; (B) Validation set.

Discussion

Radiation lung injury is caused by radiation exposure to normal lung tissue after radiotherapy for malignant tumors such as esophageal cancer in the chest.6,14 Radiation pneumonitis is an early form of radiation-induced lung injury.15 Mild cases can recover on their own, but severe cases can progress to pulmonary fibrosis, respiratory dysfunction, and even breathing difficulties. Therefore, accurately predicting the risk of radiation pneumonitis after radiotherapy can guide early intervention for patients and reduce lung damage caused by radiation pneumonitis. However, due to the non-specific early symptoms of radiation pneumonitis compared to other pneumonitis, and the delayed imaging changes in lung CT compared to clinical manifestations by 7 to 10 days, there is still a lack of effective indicators for predicting radiation pneumonitis in esophageal cancer after radiotherapy in clinical practice.16–18 In this study, we used radiomics methods to extract radiomics features of the patient’s affected lung and combined them with lung dosimetry parameters to construct a machine learning algorithm based prediction model for radiation-induced pneumonitis in esophageal cancer after radiotherapy, which is helpful for clinical decision-making.

Our study shows that the incidence of radiation pneumonitis in esophageal cancer patients is 23.68%, which is consistent with previous research reports.19–21 Our study also found that radiation pneumonitis is related to clinical staging, radiation therapy mode, and radiation dose, especially in patients with esophageal cancer. The more severe the condition, the higher the radiation dose, and the more severe the degree of radiation pneumonitis. This is because curative radiation therapy is often used for patients with advanced esophageal cancer, and the radiation dose is higher than adjuvant and palliative treatment. At the same time, the expansion of the diseased tissue leads to an increase in the target area of radiation therapy and damage to normal lung tissue, resulting in acute radiation injury.

In previous studies, dose volume parameters were often used as the main evaluation factor for radiation therapy.22–24 In this study, esophageal PTVD90, esophageal PTVV50, total dose, MLD, V5, V10, V15, V20 were associated with the occurrence of radiation pneumonitis and were independent risk factors for radiation pneumonitis. This is closely related to the increase in esophageal PTVD90 radiation dose, the larger the PTVV50 exposure area, the corresponding radiation dose and proportion of lung tissue, and the more severe the radiation defense.25 This suggests that dose volume parameters are closely related to the occurrence of radiation pneumonitis, and low dose volume parameters have certain value in predicting radiation pneumonitis. Therefore, controlling V5, V10, V15, V20, and MLD is a key element in reducing the occurrence of radiation pneumonitis.

The high dimensionality of feature data is one of the huge challenges faced by machine learning.26,27 Eliminating a large amount of redundant feature data while retaining key feature information can reduce the training time of the constructed machine learning model and produce good classification performance.28–30 Our study adopts a machine learning based feature selection method, ultimately selecting and retaining 8 optimal feature combinations, including one dosimetric feature, two first-order features based on wavelet decomposition calculation, and five texture features. Especially, texture features play a very important role in imaging omics research, providing information about organizational structure and spatial distribution. It is encouraging that this study, based on imaging omics and dosimetry models, can effectively predict the risk of radiation pneumonitis after esophageal cancer radiotherapy and has achieved good prediction results in both the training and testing sets, especially reaching an AUC value of 0.891 in the training set. This indicates that the machine learning model has a high net profit and risk threshold range, which also indicates that the radiation pneumonitis prediction model we constructed has high reliability and practicality.

Our study also has some inevitable limitations. Firstly, due to the small sample size of this study, it may have an impact on the stability of the model. Therefore, in the future, large sample cohort studies are still needed to verify the predictive model; Secondly, this study is a single-center study, and all patient data are from the same simulated positioning CT. Therefore, further research and verification are needed to determine whether similar predictive effects can be achieved on different positioning CT data. In view of this, we plan to incorporate multi-center data and analyze CT data from different locations in future research, continuously optimizing the model to improve its generalization ability. Nonetheless, we cannot deny that the model that combines pre radiotherapy localization CT imaging features with diseased lung dosimetry parameters to predict the occurrence of radiation-induced pneumonitis in esophageal cancer after radiotherapy has certain clinical practical value. The radiation-induced pneumonitis prediction model constructed based on machine learning algorithms can assist doctors in identifying high-risk patients early, providing reference and formulating personalized prevention and treatment strategies to improve treatment effectiveness and reduce adverse reactions, thereby effectively improving patient prognosis.

Conclusion

In conclusion, radiation pneumonitis occurs in patients with esophageal cancer, which is related to the severity of the condition and the parameters of radiation dose volume. The combination of radiomics features and dosimetric parameters of the affected lung has good predictive value for radiation pneumonitis in esophageal cancer patients after radiotherapy. Especially with the combination of imaging and lung dosimetry parameters, the random forest prediction model has the optimal predictive performance, which can help medical personnel develop personalized prevention and intervention strategies.

Disclosure

The authors report no conflicts of interest in this work.

References

1. Uhlenhopp DJ, Then EO, Sunkara T, Gaduputi V. Epidemiology of esophageal cancer: update in global trends, etiology and risk factors. Clin J Gastroenterol. 2020;13(6):1010–1021. doi:10.1007/s12328-020-01237-x

2. Watanabe M, Otake R, Kozuki R, et al. Recent progress in multidisciplinary treatment for patients with esophageal cancer. Surg Today. 2020;50(1):12–20. doi:10.1007/s00595-019-01878-7

3. Yao Y, Lu J, Qin Z, et al. High-dose versus standard-dose radiotherapy in concurrent chemoradiotherapy for inoperable esophageal cancer: a systematic review and meta-analysis. Radiother Oncol. 2023;184:109700. doi:10.1016/j.radonc.2023.109700

4. Bledsoe TJ, Nath SK, Decker RH. Radiation pneumonitis. Clinics Chest Med. 2017;38(2):201–208. doi:10.1016/j.ccm.2016.12.004

5. Sun L, Wang Y, Zhu L, et al. Analysis of the risk factors of radiation pneumonitis in patients after radiotherapy for esophageal squamous cell carcinoma. Front Oncol. 2023;13:1198872. doi:10.3389/fonc.2023.1198872

6. Hanania AN, Mainwaring W, Ghebre YT, Hanania NA, Ludwig M. Radiation-induced lung injury: assessment and management. Chest. 2019;156(1):150–162. doi:10.1016/j.chest.2019.03.033

7. Ullah T, Patel H, Pena GM, Shah R, Fein AM. A contemporary review of radiation pneumonitis. Curr Opin Pulm Med. 2020;26(4):321–325. doi:10.1097/MCP.0000000000000682

8. Käsmann L, Dietrich A, Staab-Weijnitz CA, et al. Radiation-induced lung toxicity - cellular and molecular mechanisms of pathogenesis, management, and literature review. Radiat Oncol. 2020;15(1):214. doi:10.1186/s13014-020-01654-9

9. Zhang Z, Wang Z, Yan M, et al. Radiomics and dosiomics signature from whole lung predicts radiation pneumonitis: a model development study with prospective external validation and decision-curve analysis. Int J Radiat Oncol Biol Phys. 2023;115(3):746–758. doi:10.1016/j.ijrobp.2022.08.047

10. Kraus KM, Oreshko M, Schnabel JA, Bernhardt D, Combs SE, Peeken JC. Dosiomics and radiomics-based prediction of pneumonitis after radiotherapy and immune checkpoint inhibition: the relevance of fractionation. Lung Cancer. 2024;189:107507. doi:10.1016/j.lungcan.2024.107507

11. Zhou L, Wen Y, Zhang G, Wang L, Wu S, Zhang S. Machine learning-based multiomics prediction model for radiation pneumonitis. J Oncol. 2023;2023:5328927. doi:10.1155/2023/5328927

12. Chen Q, Zhang L, Liu S, et al. Radiomics in precision medicine for gastric cancer: opportunities and challenges. Eur Radiol. 2022;32(9):5852–5868. doi:10.1007/s00330-022-08704-8

13. Deo RC. Machine learning in medicine. Circulation. 2015;132(20):1920–1930. doi:10.1161/CIRCULATIONAHA.115.001593

14. Arroyo-Hernández M, Maldonado F, Lozano-Ruiz F, Muñoz-Montaño W, Nuñez-Baez M, Arrieta O. Radiation-induced lung injury: current evidence. BMC Pulm Med. 2021;21(1):9. doi:10.1186/s12890-020-01376-4

15. Roy S, Salerno KE, Citrin DE. Biology of radiation-induced lung injury. Sem Rad Oncol. 2021;31(2):155–161. doi:10.1016/j.semradonc.2020.11.006

16. Liang B, Tian Y, Chen X, et al. Prediction of radiation pneumonitis with dose distribution: a convolutional neural network (CNN) based model. Front Oncol. 2019;9:1500. doi:10.3389/fonc.2019.01500

17. Yakar M, Etiz D, Metintas M, Ak G, Celik O. Prediction of radiation pneumonitis with machine learning in stage III lung cancer: a pilot study. Technol Cancer Res Treat. 2021;20:15330338211016373. doi:10.1177/15330338211016373

18. Jiang W, Song Y, Sun Z, Qiu J, Shi L. Dosimetric factors and radiomics features within different regions of interest in planning CT images for improving the prediction of radiation pneumonitis. Int J Radiat Oncol Biol Phys. 2021;110(4):1161–1170. doi:10.1016/j.ijrobp.2021.01.049

19. Marks LB, Bentzen SM, Deasy JO, et al. Radiation dose-volume effects in the lung. Int J Radiat Oncol Biol Phys. 2010;76(3 Suppl):S70–6. doi:10.1016/j.ijrobp.2009.06.091

20. Mehta V. Radiation pneumonitis and pulmonary fibrosis in non-small-cell lung cancer: pulmonary function, prediction, and prevention. Int J Radiat Oncol Biol Phys. 2005;63(1):5–24. doi:10.1016/j.ijrobp.2005.03.047

21. Tonison JJ, Fischer SG, Viehrig M, et al. Radiation pneumonitis after intensity-modulated radiotherapy for esophageal cancer: institutional data and a systematic review. Sci Rep. 2019;9(1):2255. doi:10.1038/s41598-018-38414-5

22. Uchida Y, Tsugawa T, Tanaka-Mizuno S, et al. Prediction of radiation pneumonitis using dose-volume histogram parameters with high attenuation in two types of cancer: a retrospective study. PLoS One. 2020;15(12):e0244143. doi:10.1371/journal.pone.0244143

23. Piotrowski T, Matecka-Nowak M, Milecki P. Prediction of radiation pneumonitis: dose-volume histogram analysis in 62 patients with non-small cell lung cancer after three-dimensional conformal radiotherapy. Neoplasma. 2005;52(1):56–62.

24. Kim TH, Cho KH, Pyo HR, et al. Dose-volumetric parameters for predicting severe radiation pneumonitis after three-dimensional conformal radiation therapy for lung cancer. Radiology. 2005;235(1):208–215. doi:10.1148/radiol.2351040248

25. Huang Y, Feng A, Lin Y, et al. Radiation pneumonitis prediction after stereotactic body radiation therapy based on 3D dose distribution: dosiomics and/or deep learning-based radiomics features. Radiat Oncol. 2022;17(1):188. doi:10.1186/s13014-022-02154-8

26. Haug CJ, Drazen JM. Artificial intelligence and machine learning in clinical medicine, 2023. New Engl J Med. 2023;388(13):1201–1208. doi:10.1056/NEJMra2302038

27. Hong GS, Jang M, Kyung S, et al. Overcoming the challenges in the development and implementation of artificial intelligence in radiology: a comprehensive review of solutions beyond supervised learning. Korean J Radiol. 2023;24(11):1061–1080. doi:10.3348/kjr.2023.0393

28. MacEachern SJ, Forkert ND. Machine learning for precision medicine. Genome. 2021;64(4):416–425. doi:10.1139/gen-2020-0131

29. Krishnan R, Rajpurkar P, Topol EJ. Self-supervised learning in medicine and healthcare. Nat Biomed Eng. 2022;6(12):1346–1352. doi:10.1038/s41551-022-00914-1

30. An Q, Rahman S, Zhou J, Kang JJ. A comprehensive review on machine learning in healthcare industry: classification, restrictions, opportunities and challenges. Sensors. 2023;23(9):4178. doi:10.3390/s23094178