Computed tomography-based deep learning prediction of radiation-induced functional decline after lung cancer radiotherapy

Introduction

Pulmonary function decline is a major concern following thoracic radiotherapy, particularly in patients with lung cancer who often have limited respiratory reserves (1). Spirometry indices such as forced vital capacity (FVC), forced expiratory volume in 1 s (FEV1), and FEV1/FVC ratio reflect lung volume, airflow limitation, and overall respiratory status, with values of <70% commonly indicating clinically significant obstruction (2). Changes in these parameters after radiotherapy may represent both treatment-related injury and underlying pulmonary vulnerability, emphasizing the importance of predicting post-treatment functional outcomes (3,4).

Radiotherapy is used in approximately 50–60% of patients with lung cancer; however, treatment-related pulmonary injury remains a significant challenge despite advances in delivery techniques (5). Previous studies have reported measurable declines in pulmonary function after radiotherapy, including reductions in FVC and FEV1 and progressive deterioration over time (6). Disproportionate decreases in FEV1 relative to FVC may lead to reduced FEV1/FVC ratios, reflecting clinically meaningful pulmonary functional deterioration (7).

Pulmonary function decline is associated with reduced quality of life and poorer survival outcomes, and lower FEV1, FVC, and FEV1/FVC values have been linked to increased mortality risk (8,9). Importantly, some patients with normal baseline function develop significant impairment after treatment, highlighting the need for early risk stratification (10,11). Current risk assessments rely primarily on baseline pulmonary function testing and dose-volume parameters, such as mean lung dose and V20; however, accumulating evidence suggests that these parameters alone provide limited predictive accuracy for individual patient outcomes (12,13). Pretreatment chest computed tomography (CT) scans contain structural information correlated with lung function, and deep learning enables the automated extraction of complex imaging features; however, limited datasets and class imbalance remain challenges (14,15).

In this study, we aimed to develop and validate a deep learning-based model to predict post-radiation pulmonary function using pre-treatment CT and clinical parameters. A three-dimensional convolutional neural network integrating imaging features and clinical variables was designed to estimate post-treatment FVC, FEV1, and FEV1/FVC and to identify patients at risk of transitioning to a high-risk functional state. Through this framework, we sought to enable the early prediction of treatment-induced functional decline and support individualized decision-making during radiotherapy planning. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0286/rc).

Methods

Study population

This retrospective study included patients with stage I–II lung cancer who received definitive radiotherapy at Ajou University Hospital in Suwon, Republic of Korea, from January 2010 to October 2025. The inclusion criteria were as follows: (I) age ≥18 years; (II) availability of pre-treatment chest CT scans; (III) spirometry performed both before and after radiotherapy; and (IV) complete radiotherapy parameter data. Patients who underwent lung resection, did not complete radiotherapy, or had inadequate image quality were excluded (Figure 1). In total, 276 patients were included in this study. The cohort was partitioned at the patient level using five-fold cross-validation. In each fold, approximately 220–221 patients (~80%) were used for training and 55–56 patients (~20%) for validation, with no patient overlap between training and validation sets. Aggregated across the five folds, each patient contributed exactly once to a validation set. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Ajou University Hospital (IRB No. AJOUIRB-DB-2025-635), and the requirement for informed consent was waived because of the retrospective design.

Figure 1 Flowchart of the patient selection process. A total of 305 patients were initially assessed, and 276 patients with stage I–II lung cancer were finally included after applying exclusion criteria such as lack of pre-treatment CT, absence of follow-up spirometry, or history of lung resection. CT, computed tomography; FEV1, forced expiratory volume in 1 s; FVC, forced vital capacity.

Imaging and clinical data

Chest CT images acquired from multiple scanners were collected prior to radiotherapy. For preprocessing, images were converted from DICOM to NIfTI format using dicom2nifti and resampled to 2.0 mm isotropic resolution using nibabel and SciPy. Lung segmentation was performed using TotalSegmentator, a publicly available pretrained nnU-Net-based model for multi-organ segmentation in CT images (16). The segmented lung volumes were then intensity-scaled, resized to 128×128×128 voxels, and saved as tensors using MONAI and PyTorch for downstream deep learning analysis. Clinical variables included baseline pulmonary function parameters (FVC, FEV1, and FEV1/FVC), all of which were obtained prior to radiotherapy. Pulmonary function tests were performed using a closed-circuit spirometer (MedGraphics CPFS/D BREEZE; Medical Graphics Corp., St Paul, MN, USA) according to standard spirometry guidelines. In addition, treatment-related variables, including total radiation dose and number of fractions, were collected from radiotherapy records. These predictors were integrated into the prediction model.

Deep learning model

The overall proposed model was designed as a multimodal architecture integrating three-dimensional (3D) CT imaging features and clinical variables (Figure 2). A pretrained 3D ResNet-18 encoder was used to automatically extract imaging features from the lung region segmented from pre-treatment chest CT scans. The encoder was initialized with MedicalNet weights pretrained on the 3D Seg-8 dataset, which comprises diverse modalities, target organs, and pathologies (17), and was kept frozen during training to mitigate overfitting given the limited dataset size. The extracted 512-dimensional imaging feature vector was concatenated with five standardized clinical variables: baseline FVC, baseline FEV1, baseline FEV1/FVC, radiation dose, and number of fractions. The combined feature vector was passed through fully connected layers with rectified linear unit activation and dropout to predict the post-treatment FVC, FEV1, and FEV1/FVC. The training objective consisted of a weighted mean squared error for the three outcomes, with weights of 0.35, 0.35, and 0.30 assigned to FVC, FEV1, and FEV1/FVC, respectively. To emphasize clinically relevant deterioration, a focal-weighted FEV1/FVC loss component was added with a weight of 0.3; within this component, errors from high-risk samples, defined by a low post-treatment FEV1/FVC or a marked decline from baseline, were multiplied by 5.0. Model optimization was performed using the AdamW optimizer. Five-fold cross-validation was applied, and ensemble predictions were obtained by averaging the fold outputs. The model was implemented in Python 3.13 with PyTorch 2.6.0 (CUDA 12.4), and training and inference were performed on an NVIDIA GeForce RTX 3080 Ti GPU.

Figure 2 Overview of the proposed multimodal deep learning framework. The model extracts 512-dimensional imaging features from pre-treatment 3D computed tomography scans using a ResNet-18 backbone and integrates them with five clinical variables (age, sex, baseline function, and radiation dose parameters) to predict post-treatment pulmonary function indices. Imaging features were combined with clinical variables and input into fully connected layers to predict post-treatment pulmonary function (FVC, FEV1, FEV1/FVC). Five-fold cross-validation and ensemble prediction were applied. 3D, three-dimensional; CT, computed tomography; FEV1, forced expiratory volume in 1 s; FVC, forced vital capacity.

Statistical analysis

The primary outcome was post-treatment pulmonary function (FVC, FEV1, and FEV1/FVC), and high-risk functional decline was defined as a post-treatment FEV1/FVC of <70% (18). A single regression model was trained to directly predict all three pulmonary function outcomes; no separate classification model was developed. For the classification analysis, patients with a baseline FEV1/FVC of ≥70% were selected, and high-risk transition was determined by applying the same 70% threshold to the model-predicted post-treatment FEV1/FVC. Regression performance was evaluated using the mean absolute error (MAE) and concordance correlation coefficient (CCC), and classification performance using accuracy, sensitivity, specificity, and area under the curve (AUC). Performance metrics were averaged across five replicates, and 95% confidence intervals (CIs) were estimated using bootstrap resampling.

Results

Patient characteristics

The baseline demographic and clinical characteristics of the 276 patients are summarized in Table 1. The study population had a mean age of 68.51±9.98 years, with a predominance of male patients (n=229, 83.0%). Regarding radiotherapy parameters, the mean total radiation dose was 54.56±12.52 Gy, delivered in an average of 21.87±11.22 fractions.

At baseline, 129 (46.7%) patients exhibited normal pulmonary function with a pre-treatment FEV1/FVC of ≥70%, while 147 (53.3%) showed impaired function with a ratio of <70%. Analysis of post-treatment functional changes revealed that of the 129 patients with normal baseline function, 22 (17.1%) transitioned to the high-risk group (FEV1/FVC <70%) after radiotherapy, whereas 107 (82.9%) maintained their function within the normal range.

Prediction of post-treatment pulmonary function

The deep learning model demonstrated stable performance across five-fold cross-validation. In the regression analysis, the ensemble model achieved a CCC of 0.647 (95% CI: 0.622–0.673) for FVC, 0.716 (95% CI: 0.693–0.739) for FEV1, and 0.825 (95% CI: 0.807–0.843) for FEV1/FVC, indicating the highest agreement for FEV1/FVC prediction (Figure 3). The MAE was 0.419 L (95% CI: 0.402–0.436) for FVC, 0.263 L (95% CI: 0.252–0.274) for FEV1, and 5.694% (95% CI: 5.473–5.916%) for FEV1/FVC. Compared with individual-fold models, ensemble averaging improved prediction stability and yielded consistent performance across pulmonary function metrics.

Figure 3 Predicted versus observed post-treatment pulmonary function. Scatter plots (A-C) compare predicted and measured values for FVC, FEV1, and FEV1/FVC. The dashed diagonal lines indicate perfect agreement, and concordance correlation coefficients are shown. Dotted reference lines in (C) mark the FEV1/FVC threshold of 70%. Bland-Altman plots (D-F) display the agreement between predicted and measured values with mean bias and 95% limits of agreement. CCC, concordance correlation coefficient; FEV1, forced expiratory volume in 1 s; FVC, forced vital capacity; MAE, mean absolute error.

Scatter plots demonstrated strong correlations between the predicted and observed post-treatment pulmonary function values. Notably, prediction accuracy remained stable around the clinically important FEV1/FVC threshold of 70%, indicating reliable performance near the diagnostic decision boundary (Figure 3C). Bland-Altman analysis (Figure 3D-3F) showed good agreement between predicted and measured values, with near-zero bias (FVC: −0.002 L; FEV1: −0.007 L) and a small, clinically negligible bias for FEV1/FVC (−2.6%). The 95% limits of agreement were narrow relative to the clinical range of each parameter (FVC: −1.05 to 1.05 L; FEV1: −0.67 to 0.66 L; FEV1/FVC: −15.5% to 10.3%). Residuals were symmetrically distributed around the line of equality without a discernible proportional trend, supporting the robustness and stability of the model across the functional spectrum.

Prediction of high-risk functional decline

We evaluated the classification performance for predicting the transition of patients with normal pre-treatment lung function (FEV1/FVC ≥70%) to a high-risk functional state (FEV1/FVC <70%) after treatment. The predicted post-treatment FEV1/FVC values were used to assign high-risk labels at the 70% threshold and to compute the AUC as a continuous risk score. Through five-fold cross-validation, 129 of 276 (46.7%) patients were identified as having normal baseline function and were included in the analysis. The average performance across individual folds yielded a sensitivity of 73.0% and an AUC of 0.802 (Figure 4). The ensemble model, which averaged predictions from the five-fold models, achieved an accuracy of 78.1% (95% CI: 75.0–81.3%), a sensitivity of 54.5% (95% CI: 45.0–64.1%), a specificity of 83.0% (95% CI: 79.9–86.2%), and an AUC of 0.802 (95% CI: 0.762–0.842). Confusion matrix analysis revealed 89 true negatives, 18 false positives, 10 false negatives, and 12 true positives. Compared with the individual folds, the ensemble model showed an 18.5 percentage point decrease in sensitivity; however, it maintained a consistent AUC, ensuring the overall reliability and consistency of the classification performance.

Figure 4 Performance of the ensemble model in predicting high-risk transition. (A) ROC curve showing an area under the curve of 0.802. (B) Confusion matrix displaying the classification results for 129 patients with normal baseline function. AUC, area under the curve; ROC, receiver operating characteristic.

Discussion

This study developed and validated a deep learning framework that predicts radiation-induced pulmonary functional decline using pre-treatment chest CT and clinical parameters in patients with lung cancer. Importantly, the primary objective of this model was not only to estimate post-treatment spirometric values but also to provide risk stratification before radiotherapy planning begins. Unlike conventional dose-volume histogram-based metrics, which become available only after a treatment plan has been generated, the proposed approach enables the assessment of intrinsic pulmonary vulnerability at the pretreatment stage (19,20). By identifying patients susceptible to transitioning from normal baseline function to a high-risk pulmonary functional state (FEV1/FVC <70%) before dose distribution design, this framework may support earlier and more individualized therapeutic decision making. Such pre-planning risk information may inform considerations, including treatment modality selection, fractionation strategy, or proactive implementation of functional lung avoidance techniques. In this context, the model should be viewed not as a replacement for dose-based planning metrics but as a complementary tool that introduces biological susceptibility into the radiotherapy decision pathway.

A key conceptual advance of this study is the transition from cross-sectional functional estimation to longitudinal prediction of treatment-induced physiological deterioration. Previous deep learning studies predicting pulmonary function from CT imaging have largely focused on cross-sectional estimation of concurrently measured lung function, predominantly in health screening populations (14,15). Large-scale studies using low-dose CT have demonstrated high accuracy in estimating FVC and FEV1 and meaningful performance in predicting reduced FEV1/FVC ratios (21). Other machine learning approaches have similarly reported quantitative correlations between CT-derived features and pulmonary function indices (22). Although these studies established the feasibility of imaging-based functional assessments, they primarily addressed static functional estimations rather than treatment-induced changes. In contrast, the present study adopted a longitudinal clinical perspective by predicting future functional deterioration associated with radiotherapy. By shifting the predictive target from static measurement to future outcomes, this study expands the role of imaging-based artificial intelligence from descriptive assessment to prospective clinical decision making.

Notably, the model demonstrated the ability to identify patients with normal baseline pulmonary function who subsequently transitioned to a high-risk pulmonary functional state (FEV1/FVC <70%) after treatment. This subgroup is clinically important because deterioration in these patients may otherwise remain unanticipated when relying solely on baseline spirometry or conventional dose-based parameters. In this subgroup, the model achieved an AUC of 0.802 with high specificity (83.0%) and moderate sensitivity (54.5%) at the 70% FEV1/FVC threshold. Because only 17.1% of patients with preserved baseline function transitioned to a high-risk state, the decision boundary at this fixed threshold favored specificity, reflecting the relative scarcity of positive cases. In a pre-planning context this profile is informative: the high specificity helps avoid unnecessary intensified monitoring in patients likely to maintain preserved function, while the continuous risk score—supported by the preserved AUC—provides a discriminative basis for ranking individual vulnerability beyond a single fixed cutoff. The identification of an otherwise hidden vulnerable population suggests that imaging-derived biological susceptibility may complement traditional dosimetric risk assessments and improve individualized patient stratification.

Beneath this stratification capability lies a regression performance profile worth a closer look. Among the three pulmonary function outcomes, the model achieved the strongest agreement for FEV1/FVC, followed by FEV1 and FVC. This likely reflects the nature of the outcome itself: FEV1/FVC is an intrinsically normalized ratio whose distribution is more comparable across individuals than absolute lung volumes, which vary substantially with body size, age, and sex. The focal-weighted loss component additionally directed the model toward the clinically meaningful end of the FEV1/FVC distribution, further supporting performance in this range (23). Although the absolute MAE for FEV1/FVC (5.69%) appeared larger than that of FVC (0.42 L) or FEV1 (0.26 L), MAE is scale-dependent across metrics expressed in different units. When expressed relative to the cohort mean, the relative error was smallest for FEV1/FVC (~8.5%) compared with FVC (~15.1%) and FEV1 (~14.4%), consistent with the highest CCC observed for FEV1/FVC. Bland-Altman analysis additionally showed near-zero bias and limits of agreement that were narrow relative to the clinical range of each parameter, indicating the absence of systematic over- or under-prediction across the measurement range. Together, these regression-level patterns indicate that the model produces internally consistent predictions across the three spirometric outcomes, supporting its overall predictive validity at the regression level.

From a clinical standpoint, the ability to predict post-radiation pulmonary decline before treatment initiation has important implications. Early identification of high-risk patients may allow the modification of radiotherapy strategies, including dose adjustment, fractionation schemes, and treatment volumes, as well as consideration of alternative modalities. Early pulmonary rehabilitation, closer functional monitoring, and preventive interventions should be considered for vulnerable individuals.

Despite these strengths, this study has some limitations that must be acknowledged. First, it was a single-institution retrospective study that included 276 patients, which may have limited generalizability. While previous CT-based pulmonary function prediction studies included tens of thousands of screening participants, our cohort was restricted to patients undergoing radiotherapy for lung cancer, representing a smaller but more clinically complex population. Therefore, external validation using multicenter datasets is required (21). Second, the sensitivity for predicting high-risk transitions was moderate (54.5%), indicating that some at-risk patients may not have been identified. This limitation was likely influenced by class imbalance, as only 22 of 129 (17.1%) patients with normal baseline function transitioned to a high-risk status. Variability in sensitivity across folds further reflects distributional imbalances. Future refinement using advanced imbalance-handling techniques may improve the detection performance. Third, preprocessing involved resizing CT volumes to 128×128×128 voxels, which may have reduced fine-grained structural information. Higher-resolution input or attention-based architectures could potentially enhance sensitivity to localized parenchymal vulnerability. Fourth, although patients undergoing lung resection were excluded, several clinical factors that may modulate radiation-induced pulmonary changes were not systematically captured in our retrospective database, including detailed smoking exposure (e.g., pack-years and smoking duration), formal diagnoses of pre-existing obstructive lung disease, and concurrent use of inhaled or oral corticosteroids during radiotherapy. Although baseline spirometric measurements likely reflect the functional consequences of these factors to some extent, prospective collection of such information would allow more granular adjustment in future studies. Furthermore, although radiation-induced lung injury is often associated with restrictive physiological impairment, the present study focused on clinically meaningful spirometric functional deterioration using routinely available pulmonary function parameters rather than distinguishing restrictive versus obstructive pathophysiological mechanisms.

In this cohort, 17.1% of the patients with normal baseline pulmonary function transitioned to a high-risk state after radiotherapy. The identification of this subgroup highlights the clinical relevance of pre-treatment imaging-based risk stratification. The relatively high specificity observed in this study suggests that patients predicted to maintain preserved function may avoid unnecessary intensified monitoring. Beyond lung cancer radiotherapy, the focal-weighted predictive strategy may also be applicable to other medical imaging problems in which clinically critical outcomes are relatively rare but important.

Conclusions

In this study, we developed and validated a deep learning model that integrates pretreatment chest CT imaging and clinical variables to predict post-radiotherapy pulmonary functional outcomes in patients with lung cancer. The model demonstrated the ability to estimate post-treatment spirometric indices and to identify patients at risk of functional deterioration, including those transitioning from a normal baseline pulmonary function to a high-risk pulmonary functional state.

These findings suggest that imaging-based predictive modeling using pre-treatment data may provide additional information for individualized patient risk assessment before radiotherapy. Such information may complement conventional dose-based planning parameters and potentially support personalized treatment strategies and follow-up planning. Further validation using multicenter datasets and prospective studies is required to confirm the generalizability and clinical utility of this approach.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0286/rc

Data Sharing Statement: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0286/dss

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0286/prf

Funding: This research was supported by grants from the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI) (Nos. RS-2021-KH113822 and RS-2022-KH130307) and the National R&D Program for Cancer Control through the National Cancer Center (NCC) (No. RS-2025-02214710), both funded by the Ministry of Health & Welfare, Republic of Korea.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0286/coif). All authors report funding from Korea Health Industry Development Institute (KHIDI) (Nos. RS-2021-KH113822 and RS-2022-KH130307) and National Cancer Center (NCC) (No. RS-2025-02214710), both funded by the Ministry of Health & Welfare, Republic of Korea. The authors have no other 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. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Ajou University Hospital (IRB No. AJOUIRB-DB-2025-635), and the requirement for informed consent was waived because of the retrospective design.

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. Wang S, Xu D, Xiao L, et al. Radiation-induced lung injury: from mechanism to prognosis and drug therapy. Radiat Oncol 2025;20:39. [Crossref] [PubMed]
2. Agusti A, Vogelmeier CF. GOLD 2024: a brief overview of key changes. J Bras Pneumol 2023;49:e20230369. [Crossref] [PubMed]
3. Bensenane R, Helfre S, Cao K, et al. Optimizing lung cancer radiation therapy: A systematic review of multifactorial risk assessment for radiation-induced lung toxicity. Cancer Treat Rev 2024;124:102684. [Crossref] [PubMed]
4. Lee TH, Kang BH, Kim HJ, et al. Predictors of Post-chemoradiotherapy Pulmonary Complication in Locally Advanced Non-Small Cell Lung Cancer. Cancer Res Treat 2023;55:865-74. [Crossref] [PubMed]
5. Corral J, Borras JM, Lievens Y. Utilisation of radiotherapy in lung cancer: A scoping narrative literature review with a focus on the introduction of evidence-based therapeutic approaches in Europe. Clin Transl Radiat Oncol 2024;45:100717. [Crossref] [PubMed]
6. Tanaka H, Hironaka A, Nakabe C, et al. Impact of Lung Dose on Pulmonary Function Decline after Stereotactic Body Radiation Therapy for Lung Cancer. International Journal of Radiation Oncology*Biology*Physics 2025;123:e216-7.
7. Backman H, Vanfleteren LEGW, Mannino DM, et al. Severity of Airflow Obstruction Based on FEV1/FVC Versus FEV1Percent Predicted in the General U.S. Population. American Journal of Respiratory and Critical Care Medicine 2024;210:1308-16.
8. Shen YL, Shen CI, Chiang CL, et al. Reduced FEV(1) as Prognostic Factors in Patients With Advanced NSCLC Receiving Immune Checkpoint Inhibitors. Front Med (Lausanne) 2022;9:860733. [Crossref] [PubMed]
9. Costanzo S, Magnacca S, Bonaccio M, et al. Reduced pulmonary function, low-grade inflammation and increased risk of total and cardiovascular mortality in a general adult population: Prospective results from the Moli-sani study. Respir Med 2021;184:106441. [Crossref] [PubMed]
10. Yan Y, Zhu Y, Yang S, et al. Clinical predictors of severe radiation pneumonitis in patients undergoing thoracic radiotherapy for lung cancer. Transl Lung Cancer Res 2024;13:1069-83. [Crossref] [PubMed]
11. Badola A, Gupta M, Bansal S, et al. The predictive role of baseline pulmonary function test in lung carcinoma patients for radiation-induced lung toxicity treated with conformal radiation therapy. Indian J Cancer 2024;61:75-80. [Crossref] [PubMed]
12. Zhang Y, Zhou SH, Yan YJ, et al. Predicting Severe Radiation Pneumonitis in Patients With Locally-Advanced Non-Small-Cell Lung Cancer After Thoracic Radiotherapy: Development and Validation of a Nomogram Based on the Clinical, Hematological, and Dose-Volume Histogram Parameters. Clin Lung Cancer 2025;26:393-406. [Crossref] [PubMed]
13. Zhu Y, Yan Y, Fu J, et al. Reducing radiation pneumonitis in lung cancer patients: from Chinese expert consensus to practice. Transl Lung Cancer Res 2025;14:2747-59. [Crossref] [PubMed]
14. Wang W, Sun Y, Wu R, et al. Machine-Learning-Based Computed Tomography Radiomics Regression Model for Predicting Pulmonary Function. Acad Radiol 2025;32:4278-88. [Crossref] [PubMed]
15. Park H, Yun J, Lee SM, et al. Deep Learning-based Approach to Predict Pulmonary Function at Chest CT. Radiology 2023;307:e221488. [Crossref] [PubMed]
16. Wasserthal J, Breit HC, Meyer MT, et al. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiol Artif Intell 2023;5:e230024. [Crossref] [PubMed]
17. Chen S, Ma K, Zheng Y. Med3d: Transfer learning for 3d medical image analysis. arXiv 2019. arXiv:190400625.
18. Berg A, Rokach A, Bohadana A, et al. Using the FEV1/FVC < 0.7 as Criterion for Obstruction in the Geriatric Population: Is There a Risk of Overdiagnosis? The Israel Medical Association journal IMAJ 2025;27:754-8.
19. Flakus MJ, Kent SP, Wallat EM, et al. Metrics of dose to highly ventilated lung are predictive of radiation-induced pneumonitis in lung cancer patients. Radiother Oncol 2023;182:109553. [Crossref] [PubMed]
20. Ruysscher D, Wauters E, Jendrossek V, et al. Diagnosis and treatment of radiation induced pneumonitis in patients with lung cancer: An ESTRO clinical practice guideline. Radiother Oncol 2025;207:110837. [Crossref] [PubMed]
21. Li R, Guo H, Wu Q, et al. Deep learning based CT images for lung function prediction in patients with chronic obstructive pulmonary disease. BMC Pulm Med 2025;25:484. [Crossref] [PubMed]
22. Porazzi A, Zaffaroni M, Pierini VE, et al. CT-Based Radiomics Enhance Respiratory Function Analysis for Lung SBRT. Bioengineering (Basel) 2025;12:800. [Crossref] [PubMed]
23. Yeung M, Sala E, Schönlieb CB, et al. Unified Focal loss: Generalising Dice and cross entropy-based losses to handle class imbalanced medical image segmentation. Comput Med Imaging Graph 2022;95:102026. [Crossref] [PubMed]