A transfer learning-based hybrid deep- and machine-learning regression approach for predicting the postoperative pulmonary function.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, Huadong Hospital, Fudan University, 200040, Shanghai, People's Republic of China.
- Department of Radiology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, 200127, Shanghai, People's Republic of China.
- Kashi Prefecture Second People's Hospital, 844000, Xinjiang, People's Republic of China.
- Department of Radiology, Huadong Hospital, Fudan University, 200040, Shanghai, People's Republic of China. [email protected].
- Department of Radiology, Shanghai Chest Hospital, School of Medicine, Shanghai Jiao Tong University, 200030, Shanghai, People's Republic of China. [email protected].
Abstract
Accurately predicting postoperative pulmonary function is essential for surgical decision-making in patients with pulmonary nodules. Here, we developed a hybrid model using preoperative computed tomography (CT) images and evaluated its accuracy and interpretability for predicting postoperative pulmonary function. This retrospective study included 136 patients who underwent preoperative chest CT and postoperative pulmonary function tests. A pre-trained Inflated 3D ConvNet (I3D) model was fine-tuned for transfer learning to predict the postoperative pulmonary function, and the resulting model served as the feature extractor. Extracted deep learning features were combined with an elastic net regression for prediction. Models established using clinical features or radiomics with elastic net, end-to-end I3D transfer learning, and 3D ResNet18 trained from scratch were compared, with additional evaluation against a conventional segment-counting method. Performance was evaluated against spirometry, and interpretability was assessed using Grad-CAM. The hybrid model exhibited the best external test performance. For postoperative forced vital capacity (FVC) prediction, the concordance correlation coefficient (CCC) was 0.707, the Pearson correlation coefficient (Pearson r) was 0.796, and the R-squared (R<sup>2</sup>) was 0.499. For postoperative forced expiratory volume in 1 s (FEV<sub>1</sub>) prediction, the CCC was 0.729, Pearson r was 0.772, and R<sup>2</sup> was 0.418. Grad-CAM revealed the inferior and paravertebral lung regions in the FVC model and bilateral lung bases and diaphragmatic areas in the FEV<sub>1</sub> model. The hybrid elastic net model based on fine-tuned I3D features predicted postoperative pulmonary function without requiring preoperative spirometry or detailed surgical planning and may guide the development of future predictive models. Question CT-only prediction of postoperative pulmonary function remains an unmet clinical need due to the non-routine use of pulmonary function testing. Findings A hybrid CT-based model enables prediction of postoperative pulmonary function and outperforms conventional segment-counting, clinical, radiomics, and end-to-end deep learning approaches. Relevance Statement This hybrid CT-based model can estimate postoperative pulmonary function from routine preoperative non-contrast CT images, potentially assisting surgical decision-making for lung cancer.