Related Experiment Video
Updated: Jul 12, 2026
Three-Dimensional Reconstruction for the Whole Lung with Early Multiple Pulmonary Nodules
Published on: October 13, 2023
Computed tomography-based deep learning prediction of radiation-induced functional decline after lung cancer
Jung Chun Kim1, Seonhwa Kim2, Jun Hyeong Park1
1Department of Biomedical Sciences, Graduate School of Ajou University, Suwon, Republic of Korea.
Background:
Radiation therapy is a key treatment modality for lung cancer; however, post-treatment pulmonary function decline remains a clinically significant concern. Conventional dose-volume parameters provide limited individualized prediction of functional deterioration and fail to account for the heterogeneous baseline lung status visible on computed tomography (CT). In this study, we aimed to develop and validate a deep learning model to predict post-radiation pulmonary function and identify patients transitioning to a high-risk functional state [forced expiratory volume in 1 s (FEV1) to forced vital capacity (FVC) ratio, FEV1/FVC <70%] using pre-treatment chest CT and clinical variables.
Methods:
In this retrospective, single-institution study, the data of 276 patients with stage I-II lung cancer who underwent radiotherapy and had available pre- and post-treatment spirometry results were analyzed. A pretrained three-dimensional (3D) ResNet-18 architecture was used to extract imaging features from pre-treatment CT scans, which were integrated with clinical variables including age, sex, baseline pulmonary function, and radiation dose parameters. A focal-weighted loss strategy was incorporated to address the class imbalance. The model performance was evaluated using five-fold cross-validation and ensemble prediction.
Results:
The ensemble model achieved a concordance correlation coefficient of 0.65 for FVC, 0.72 for FEV1, and 0.83 for FEV1/FVC. For classification of high-risk functional decline (post-treatment FEV1/FVC <70%) among patients with normal baseline function, the model achieved an accuracy of 78.1%, a sensitivity of 54.5%, a specificity of 83.0%, and an area under the curve of 0.802. Among the patients with normal baseline pulmonary function, 17.1% transitioned to the high-risk group after treatment.
Conclusions:
Deep learning using pre-treatment CT can predict radiation-induced functional decline and provide high-confidence risk stratification. By identifying patients who are susceptible to clinically significant pulmonary functional deterioration before treatment, this approach enables individualized adaptive planning and early preventive interventions, thereby advancing precision radiation oncology.
