A hybrid artificial intelligence framework combining transfer-learned deep neural representations with regularized regression can accurately estimate postoperative pulmonary function from routine preoperative non-contrast computed tomography scans. Developed by researchers evaluating surgical outcomes in patients with pulmonary nodules, the two-stage model directly estimates key spirometric parameters, including forced vital capacity and forced expiratory volume in one second, without depending on preoperative pulmonary function testing or elaborate surgical planning. By bridging deep learning feature extraction with classical machine learning, the system provides an automated approach to support individualized clinical decision-making across common thoracic resections.
Surgical resection remains the primary curative therapy for lung cancer, which continues to be the leading contributor to cancer-related mortality globally. With modern imaging advances, incidental pulmonary nodules are now detected in twenty to thirty-one percent of chest computed tomography examinations, driving a corresponding increase in thoracic operations. In selecting between lobectomy, segmentectomy, or wedge resection, thoracic surgeons must balance oncologic adequacy with the preservation of functional lung parenchyma.
Historically, clinicians have relied on anatomical heuristics, primarily the segment-counting method, which estimates functional loss by assuming a proportional reduction based on the number of resected lung segments relative to an anatomical total of nineteen segments across both lungs. However, segment-counting and related volumetric imaging techniques exhibit fundamental shortcomings in routine practice. These conventional approaches are generally applicable only to anatomical segmentectomies and lobectomies, making them unsuitable for wedge resections where resected lung volume is smaller than a formal segment. Furthermore, they depend strictly on baseline preoperative spirometry and frequently overestimate actual functional loss by failing to account for non-proportional, lobe-dependent compensatory recovery.
To overcome these clinical limitations, investigators designed a retrospective multicenter study utilizing patient records from two medical institutions. The development cohort was derived from Huadong Hospital affiliated to Fudan University, where an initial screening of 7,494 patients who completed pulmonary function tests between October 2014 and October 2023 yielded 105 surgical patients meeting inclusion criteria. These patients had undergone lung resection, possessed thin-slice non-contrast chest computed tomography scans, and had documented postoperative spirometry. The participants, comprising 61 females and 44 males with a mean age of 65.4 years, were randomized into training and internal test cohorts using a three-to-one ratio. An independent external test cohort was established from Heilongjiang Provincial Hospital, incorporating 31 patients treated between January 2017 and January 2023.
The central architecture of the hybrid pipeline, designated D + E, leverages a pre-trained Inflated 3D ConvNet originally trained on an independent, non-overlapping cohort to predict pulmonary function metrics from non-contrast computed tomography scans of non-surgical patients. To adapt the network for postoperative forecasting, investigators unfroze the final fifteen layers of the backbone and fine-tuned the model across fifty epochs using five-fold cross-validation on the training set. Six standardized clinical variables—patient sex, age, height, weight, resection location, and resection type—were concatenated with the penultimate 500-unit dense layer, feeding into a 64-unit dense layer. The resulting 64-dimensional out-of-fold feature representations were extracted and used to train an elastic net regression model whose regularization parameters were optimized using ten-fold cross-validation.
For rigorous benchmarking, the investigators compared the hybrid D + E framework against four alternative computational modeling strategies trained on identical data. These included a radiomics pipeline combining extracted imaging features with elastic net regression (R + E), a baseline model combining the six clinical features with elastic net regression (C + E), an end-to-end 3D ResNet18 convolutional network trained entirely from scratch (Res), and a standalone fine-tuned Inflated 3D ConvNet operating without the secondary elastic net regression stage (D). Additionally, the hybrid model was evaluated directly against conventional segment-counting calculations among qualifying patients undergoing lobectomy or segmentectomy.
In external cohort evaluations, the hybrid D + E model demonstrated the highest overall predictive accuracy and the lowest error among all evaluated pipelines. For predicting postoperative forced vital capacity, the model achieved a concordance correlation coefficient of 0.707, a Pearson correlation coefficient of 0.796, an R-squared of 0.499, a mean absolute error of 0.219 liters, and a root mean square error of 0.271 liters. For predicting postoperative forced expiratory volume in one second, the model achieved a concordance correlation coefficient of 0.729, a Pearson correlation coefficient of 0.772, an R-squared of 0.418, a mean absolute error of 0.202 liters, and a root mean square error of 0.253 liters, reflecting substantial agreement between model estimates and spirometry measurements.
Alternative modeling approaches exhibited significant performance deficiencies during independent external validation. While the clinical-feature baseline (C + E) showed moderate correlation, its concordance correlation coefficient and R-squared were substantially lower (concordance correlation of 0.411 for forced vital capacity and 0.322 for forced expiratory volume), indicating poor absolute agreement. The radiomics approach suffered a severe decline in external generalizability, with its concordance correlation coefficient dropping to -0.068 for forced expiratory volume. The end-to-end deep model without elastic net yielded poor predictive strength, and the 3D ResNet18 trained from scratch performed worst across all metrics. Furthermore, permutation and Wilcoxon signed-rank tests verified that the hybrid pipeline significantly outperformed the traditional segment-counting method in error reduction and metric agreement.
Investigators also explored the potential clinical utility of the hybrid system using a binary classification threshold set at eighty percent of predicted normal lung function, a standard clinical benchmark for functional impairment. In the independent external test cohort, the hybrid model achieved an area under the receiver operating characteristic curve of 0.745 for forced vital capacity percentage predicted and 0.758 for forced expiratory volume in one second percentage predicted. At this clinically relevant cutoff, the model demonstrated a sensitivity of 90.9 percent for identifying patients with impaired forced vital capacity and a specificity of 87.5 percent for classifying forced expiratory volume impairments.
To address the interpretability of high-dimensional deep representations, researchers employed SHapley Additive exPlanations alongside gradient-weighted class activation mapping (Grad-CAM). The attribution analysis showed that nine abstract deep features were shared among the top twenty most influential variables for both respiratory outcomes. Spatial attention maps revealed that the forced vital capacity model concentrated attention on the lower thoracic zones, particularly the paravertebral soft tissues adjacent to the T12 vertebra and adjacent lung parenchyma, aligning with physiological evidence regarding chest wall compliance and respiratory musculature. Conversely, the forced expiratory volume model focused heavily on the bilateral lung bases and diaphragmatic regions, reflecting established associations between diaphragmatic motion and dynamic expiratory flow.
Despite its demonstrated predictive capabilities, the authors noted several limitations inherent to the retrospective study design. The overall sample size across both cohorts remained relatively modest because postoperative spirometry is not performed routinely in standard clinical follow-up and is typically reserved for symptomatic or clinically indicated cases, introducing non-random data attrition. Additionally, the clinical inputs were restricted to six basic parameters. The investigators emphasized that while transfer-learned representations provide an efficient, low-computation alternative to traditional counting rules, larger prospective multicenter studies with standardized postoperative testing will be necessary before implementing the tool in clinical surgical workflows.
Subject of Research: Predicting postoperative pulmonary function using a hybrid deep-learning and machine-learning model on preoperative CT scans
Article Title: A transfer learning-based hybrid deep- and machine-learning regression approach for predicting the postoperative pulmonary function
Article References: Wang, W., Sun, Y., Ma, H., & Li, M. (2026). A transfer learning-based hybrid deep- and machine-learning regression approach for predicting the postoperative pulmonary function. European Radiology Experimental, 10(1), Article 146. https://doi.org/10.1186/s41747-026-00821-z
Image Credits: AI Generated
DOI: 10.1186/s41747-026-00821-z
Keywords: pulmonary function, computed tomography, deep learning, elastic net regression, thoracic surgery, lung cancer, transfer learning, medical imaging, transfer, learning-based, hybrid, deep-
News Source: Ophelia Keating. (November 8, 2026). Hybrid AI Model Predicts Postoperative Lung Function from Preoperative CT Scans. Scienmag.
