Chronic obstructive pulmonary disease (COPD) is a chronic respiratory disease characterized by persistent airflow limitation, primarily caused by airway and lung inflammation in response to harmful particles or gases (Yildirim et al., 2025). The global initiative for chronic obstructive lung disease (GOLD) emphasizes that COPD, with its high morbidity and disability rates, is a leading cause of death (Agustí et al., 2023). Furthermore, according to the GOLD grading criteria, as COPD progresses from GOLD stage 1 to stage 4, patients experience a gradual decline in pulmonary ventilation function and exercise tolerance, alongside a substantially increased risk of cardiovascular complications (Silva et al., 2024; Li T. et al., 2025).

Among cardiovascular diseases, unstable angina (UA) has a higher incidence tendency in patients with COPD (de Miguel-Díez et al., 2024). Their shared pathophysiological basis includes systemic inflammatory response, oxidative stress, increased sympathetic nerve activity, and endothelial dysfunction induced by hypoxia (Polman et al., 2024). This cardiopulmonary comorbidity not only exacerbates the degree of myocardial ischemia but also substantially reduces cardiopulmonary reserve capacity during acute episodes, adversely affecting prognosis (Hawkins et al., 2024). Therefore, the comprehensive management of COPD complicated by UA must simultaneously address the assessment and intervention of both respiratory and cardiovascular function.

Pulmonary rehabilitation and cardiac rehabilitation have been demonstrated to substantially improve exercise tolerance, symptoms, and quality of life in patients with chronic respiratory and cardiovascular diseases (Jenkins et al., 2024; Hua et al., 2024). For patients with COPD complicated by UA, coordinated cardiopulmonary rehabilitation (CPR) is expected to yield synergistic benefits by improving muscle metabolism, reducing ventilatory load, and optimizing cardiac hemodynamics. However, existing research predominantly focuses on individual systems, and high-quality evidence using objective imaging endpoints for combined rehabilitation interventions in this complex comorbid population remains scarce (Al-Ajlouni et al., 2024). Furthermore, while traditional functional assessments are valuable, their sensitivity to key pathological changes such as lung parenchymal and cardiac structural/perfusion alterations is limited, making it difficult to comprehensively reveal the interconnected changes in cardiopulmonary structure and function (Dal Negro et al., 2025).

In recent years, cardiothoracic high-resolution CT (HRCT) has enabled the simultaneous acquisition of information on lung parenchyma (emphysema, airways), pulmonary vasculature, coronary artery calcification, and cardiac chamber/myocardial structure, providing an important tool for studying cardiopulmonary disease interactions (Rathore et al., 2025; Sui et al., 2025). Concurrently, deep learning (DL) has seen rapid progress in medical image processing (segmentation, quantification, risk prediction) and is widely applied in the automatic quantification of pulmonary nodules, emphysema distribution, airway metrics, coronary artery lesions, and ventricular geometry (Tsampras et al., 2025). While DL for segmentation and characterization in cardiothoracic CT has matured substantially, its application in evaluating rehabilitation efficacy in patients with COPD (stratified by GOLD grade) complicated by UA remains an emerging field. Key challenges include standardizing multi-phase/multi-modal data, modeling the relationship between cardiopulmonary imaging metrics and clinical outcomes, and ensuring model generalizability and interpretability across populations with varying GOLD grades and coronary artery disease burden. Recent approaches aim to enhance clinical utility through multi-center validation, lightweight networks, and interpretability modules, and attempt to integrate imaging metrics with clinical endpoints for efficacy evaluation and subgroup analysis (De Santis et al., 2025).

Based on this, this study established a DL-driven quantitative cardiothoracic CT assessment system to evaluate the impact of systematic CPR training on cardiopulmonary function in patients with COPD (stratified by GOLD grade) complicated by UA, using objective imaging endpoints. The study will enroll participants stratified by GOLD grade, automatically extract lung parenchyma, airway, pulmonary vasculature, coronary artery, and ventricular structural metrics, and compare the correlations between imaging parameters and clinical cardiopulmonary function before and after the intervention, while assessing response differences across subgroups with varying GOLD grades and coronary artery disease burden. This aims to provide quantifiable biomarkers for personalized exercise prescription, efficacy monitoring, and the management of high-risk patients with cardiopulmonary comorbidities.

Based on a hierarchical deep-learning CT model, this study systematically evaluated the comprehensive impact of cardiopulmonary rehabilitation training on cardiopulmonary structure and function in patients with COPD complicated by unstable angina. The results showed that 12 weeks of rehabilitation training not only significantly improved clinical functional indicators but also demonstrated clear cardiopulmonary structural remodeling on imaging. This study is the first, in this comorbid population, to systematically reveal the association between improvements in lung parenchyma, airways, pulmonary vasculature, and cardiac structure—assessed through automated, quantitative imaging—and clinical functional recovery, thereby providing objective imaging biomarkers for rehabilitation efficacy.

The multi-task 3D U-Net + ResNet50 model constructed in this study demonstrated excellent performance in the automatic segmentation and quantification of lung parenchyma, airways, pulmonary vasculature, and cardiac structures (achieving a maximum Dice coefficient of 0.95), substantially outperforming existing mainstream models. Its advantage lies in the use of MTL to share feature representations, thereby enhancing the recognition capability for subtle structures such as small bronchi, microvessels, and ventricular boundaries. In recent years, similar multi-task and multi-modal fusion strategies have gradually matured in COPD imaging analysis. For instance, the multi-modal feature fusion model developed by Li et al. (2025b) successfully predicted pulmonary function parameters such as FEV₁ and FVC from chest CT (Pearson r = 0.84), highlighting the importance of integrating structural and functional information. The integrated CNN backbone U-Net architecture proposed by Golkarieh et al. (2025) demonstrated exceptionally high segmentation accuracy (Dice up to 0.9495) in lung cancer CT segmentation tasks, and revealed the model’s focus on lesion areas and subtle structures through Grad-CAM visualization. Building on this, our study further applied Grad-CAM analysis and found that the multi-task 3D U-Net + ResNet50 model’s areas of focus in emphysematous regions, small bronchial walls, and microvascular networks also showed high overlap with manual annotations, thereby validating the model’s interpretability and clinical applicability in quantitative COPD imaging analysis. Furthermore, the model’s strong generalization performance in external validation aligns with the “abnormality quantification” concept proposed by Almeida et al. (2024), which utilizes self-supervised learning to identify COPD-related abnormalities from normal lung regions, enabling objective assessment of disease severity.

After CPR training, patients in, EG showed notable improvements in emphysema-related metrics (decreased LAV%, increased Perc15 and MLD), suggesting a trend toward normalization in lung parenchymal density distribution, potentially related to improved ventilatory efficiency and reduced dynamic hyperinflation due to exercise training. The underlying mechanisms were likely multifaceted. On the one hand, regular aerobic and respiratory muscle training could enhance the strength and endurance of respiratory muscles such as the diaphragm, improve ventilatory efficiency, and reduce dynamic hyperinflation. On the other hand, systemic exercise helped improve muscle metabolism and reduce resting ventilatory demand, thereby alleviating mechanical stress on lung tissue. Furthermore, the overall enhancement of cardiopulmonary function might be accompanied by a reduction in systemic inflammation and oxidative stress, positively influencing lung parenchymal structure and pulmonary vascular function. Notably, the magnitude of improvement was most pronounced in GOLD II–III patients. This finding is consistent with existing evidence regarding differential benefits of rehabilitation training across COPD severity levels, as demonstrated by He et al. (2019) clinical review and Zhang et al. (2022) RCT meta-analysis, both indicating that patients with moderate to severe disease show more substantial responses to systematic rehabilitation training in terms of exercise capacity and symptom scores. Furthermore, the reduction in WA% and Pi10 reflects the alleviation of airway remodeling, which aligns with a recent study based on an automatic metric graph neural network (AMGNN) that demonstrated improvements in airway structural metrics were substantially associated with a reduced risk of acute exacerbations (Wang et al., 2024). This study further revealed that increases in the pulmonary vascular metric BV5% and vascular fractal dimension suggest that CPR training may promote reconstruction of the pulmonary vascular network by improving endothelial function, enhancing microvascular perfusion, and optimizing ventilation/perfusion matching. Recent studies have indicated that regular exercise may upregulate endothelial nitric oxide synthase (eNOS) expression and promote nitric oxide release (Collins et al., 2025), while also modulating angiogenesis-related factors such as vascular endothelial growth factor (VEGF), thereby improving pulmonary vascular endothelial function and microcirculatory perfusion (Hong and Park, 2024). Concurrently, exercise-induced attenuation of systemic inflammation (e.g., reductions in CRP and IL-6 levels) may alleviate pulmonary vascular endothelial injury and remodeling, further optimizing ventilation/perfusion matching (Jiang et al., 2025). These mechanisms collectively may explain the observed association between pulmonary vascular structural improvements and enhanced exercise tolerance in this study. This finding is consistent with recent advances in CT radiomics for COPD phenotyping analysis (Magni et al., 2025), where BV5% has been validated as a key imaging biomarker for assessing pulmonary vascular pruning and predicting mortality risk (Cajigas et al., 2023).

This study is the first to systematically evaluate the interconnected changes in pulmonary vascular metrics (BV5% and vascular fractal dimension) and cardiac functional parameters within a rehabilitation intervention context. The notable improvements in EF and SV, along with the mild reduction in LVM observed in, EG, indicate that rehabilitation training helps optimize cardiac preload and afterload, thereby enhancing myocardial contractile efficiency. Notably, the coronary artery calcium score (CACS) showed no significant change during the 12-week intervention, which aligns with its biological nature as a marker of chronic atherosclerosis. CACS reflects the long-term cumulative burden of coronary calcification and is generally insensitive to short-term behavioral or exercise interventions; its changes require observation over longer follow-up periods (Shuval et al., 2024; Yoo et al., 2024). This aligns with the conclusions of Rathore et al. (2025) from their DL-based coronary artery calcification analysis in the NLST cohort. Correlation analysis further revealed positive associations between pulmonary vascular metrics (BV5%) and cardiopulmonary exercise capacity (VO₂max), as well as close relationships between cardiac metrics (EF, SV) and exercise tolerance (6MWD), uncovering the synergistic mechanism of the “lung-vascular-heart” axis in comorbidity rehabilitation. A recent study utilizing a graph neural network based on biphasic inspiratory and expiratory CT similarly emphasized that combined analysis of cardiopulmonary imaging features substantially improves the prediction accuracy for COPD acute exacerbations (AUC = 0.965), further supporting the clinical value of multi-system integrated assessment (Wang et al., 2024). It was worth mentioning that the biphasic (inspiratory-expiratory) CT protocol used in this study allowed for the simultaneous assessment of emphysema areas and the degree of air trapping. Expiratory-phase images, in particular, could more sensitively detect small airway disease and dynamic hyperinflation. These functional imaging parameters were closely related to the risk of COPD exacerbation, thereby significantly improving predictive accuracy compared with conventional single-phase CT.

This study also systematically validated the intrinsic relationships between quantitatively derived imaging metrics and clinical functional parameters through correlation analysis. For example, LAV% was negatively correlated with FEV₁ and 6MWD, BV5% was positively correlated with VO₂max, and EF was positively correlated with both VO₂max and 6MWD. These results are not only consistent with findings from several recent DL-based COPD imaging studies (Shen et al., 2025), but also advance the field by demonstrating that CPR not only alleviates symptoms but also induces beneficial structural remodeling. It is particularly noteworthy that the imaging biomarkers applied in this study (such as LAV%, Pi10, and BV5%) show substantial overlap with the QCT features utilized in the recently developed AutoCOPD model by Academician Zhong Nanshan’s team, indicating the core value of these metrics in screening, diagnosis, and efficacy evaluation (Lin et al., 2025). Furthermore, the Grad-CAM visualization results in this study revealed that the model’s regions of focus highly coincided with COPD pathological changes, providing strong support for the biological plausibility of the imaging biomarkers and echoing the interpretability research in airway segmentation by Garcia-Uceda et al. (2021) and Liu et al. (2025).

The core contribution of this study was that, on the basis of confirming the clinical benefits of cardiopulmonary rehabilitation, it provided, for the first time, direct and quantitative imaging evidence of beneficial cardiopulmonary structural remodeling in patients with COPD complicated by unstable angina, based on a hierarchical deep-learning imaging model. It systematically revealed the synergistic improvement mechanism of the “lung-vessel-heart” axis during rehabilitation, thereby offering a novel imaging tool and theoretical foundation for individualized rehabilitation management and efficacy monitoring in comorbid patients. However, this study had several limitations. First, as a single-center study, although patients were stratified according to GOLD classification, the relatively concentrated patient source could have introduced selection bias, which limited the generalizability of the conclusions. Second, the 12-week intervention period was relatively short. Although early improvements in some imaging and functional indicators could be observed, it was insufficient to assess changes in long-term evolution markers such as the coronary artery calcium score, nor could it confirm the durability of the cardiopulmonary structural remodeling effect. Third, the uneven distribution of sample sizes across GOLD subgroups (e.g., only 48 cases in GOLD I, compared to 156 in GOLD III) could have affected the statistical power and robustness of inter-subgroup comparisons. Finally, although the model underwent external validation, its quantitative accuracy might have been compromised in patients with very severe COPD (GOLD IV) due to severely damaged lung structure, or in patients with other cardiac comorbidities (e.g., heart failure, valvular disease) due to complex and variable cardiac structure and hemodynamics. The applicability of the model in these high-risk subgroups still required further validation. Future research could address these issues through multi-center collaboration, extended follow-up periods (e.g., ≥6 months), and expanded sample sizes to achieve a more balanced subgroup distribution. More comprehensive efficacy prediction models could be constructed by integrating multimodal data such as genomics and metabolomics. Furthermore, emerging deep learning paradigms such as hierarchical multi-instance learning and automatic metric graph neural networks could be introduced to better capture the heterogeneity of COPD, ultimately advancing precision rehabilitation management for cardiopulmonary comorbidities.

This study confirms that the DL CT image model, constructed based on multi-task 3D U-Net + ResNet50, can effectively quantify the positive impact of CPR training on cardiopulmonary structure in patients with COPD complicated by UA. This is demonstrated by reduced emphysema severity (LAV%↓), improved airway remodeling (WA%, Pi10↓), optimized pulmonary microvascular perfusion (BV5%, vascular fractal dimension↑), and enhanced cardiac function (EF, SV↑, LVM↓). These imaging improvements were substantially correlated with clinical functional metrics, providing quantifiable imaging biomarkers for evaluating rehabilitation efficacy in comorbid patients and advancing the precision management of individualized cardiopulmonary comorbidities.