Lung cancer is the second most commonly diagnosed malignancy worldwide and remains the leading cause of cancer-related deaths (1). Non–small cell lung cancer (NSCLC) accounts for approximately 80% of all lung cancer cases. Neoadjuvant immunochemotherapy (NIT), combining immune checkpoint inhibitors (ICIs) with chemotherapy prior to surgery, has emerged as a promising strategy to enhance survival outcomes in NSCLC patients (2). Accurately predicting the efficacy of NIT before surgery holds significant clinical value, as it can reduce unnecessary exposure to potentially ineffective treatments, preventing delays in administering alternative therapies. Despite the promising potential of NIT, predicting individual patient response remains a clinical challenge, as traditional radiologic assessments based on tumor size changes may not accurately reflect therapeutic efficacy (3). Although, some microscopic biomarkers such as tumor mutation burden (4) and programmed death-ligand 1 (PD-L1) expression (5) have been associated with responses to ICIs, obtaining a comprehensive evaluation of tumor status through biopsy is limited by spatial and temporal tumor heterogeneity (6). Pathologic complete response (pCR) has been recognized as a surrogate endpoint for long-term survival outcomes (2, 7–9). Notably, patients achieving pCR in trials like NADIM (8) and CheckMate 816 (2) showed significantly enhanced event-free survival. Currently, pCR status can only be determined postoperatively through histopathological examination, underscoring the need for a non-invasive, preoperative predictive method.

To address this need, non-invasive imaging techniques and advanced analytical methods have been being explored as potential solutions for preoperative prediction of pCR status. Medical imaging provides a wealth of quantitative data that can serve as potential biomarkers (10). Radiomics, the extraction of high-dimensional features from medical images, has been utilized to assess tumor diagnosis and treatment (11–14). Delta-radiomics, which analyzes the temporal changes in radiomics features across multiple time points, offers a non-invasive approach to capture treatment-induced alterations in tumors (15). Meanwhile, deep learning techniques (16, 17), particularly convolutional neural networks (CNNs), have demonstrated exceptional ability to automatically learn and extract complex patterns from medical images, surpassing traditional feature-based methods in various predictive tasks (18, 19). To date, no studies have investigated the predictive value of models that integrate delta-radiomics and deep learning features for preoperative prediction of pCR in NSCLC patients undergoing NIT.

Therefore, the aim of this study is to develop and validate a novel predictive model that combines delta-radiomics, deep learning, and clinical features to non-invasively predict pCR in NSCLC patients receiving NIT. By leveraging the complementary strengths of these approaches, we seek to provide a cost-effective and repeatable tool to aid in treatment planning and improve patient outcomes.

The current study developed a non-invasive, preoperative tool to predict pCR by integrating clinical parameters, radiomics, and deep learning features derived from CT scans. This model addresses the limitations of traditional models, which are either solely reliant on radiomics features or constrained by the learning capacity of deep learning models when used independently. By merging radiomics, clinical data, and deep learning outputs, the combined model capitalizes on the unique contributions of each data type, and effectively improves the predictive power and robustness in the non-invasive preoperative prediction of pCR in NSCLC patients receiving NIT. The incremental novelty of this work lies in being among the first to synergistically integrate Delta-radiomics with DL features for NIT response prediction. While traditional models often rely on static imaging snapshots, our approach captures both the temporal evolution of tumor heterogeneity (via Delta-radiomics) and high-level semantic morphological transformations (via DL), providing a more comprehensive spatial-temporal characterization of the tumor microenvironment during immunotherapy.

After analyzing the clinical data, serological markers did not exhibit predictive value for treatment response in this study, which contradicts some previous findings. Some studies have shown that high tumor markers before treatment are predictors of poor prognosis. Kataoka et al. (24) reported that high baseline CEA levels were associated with poor progression-free survival (PFS) in advanced NSCLC patients treated with nivolumab. Chai et al. (25) developed a nomogram to predict OS in patients with advanced NSCLC prior to ICI treatment, pointed out CYFRA21–1 level was associated with a shorter OS. Additionally, Huang et al.’s study (26) has shown that the changes in tumor markers can indicate the prognosis of patients, decreases in serological tumor markers (including CEA, CA19-9, CYFRA21-1, and NSE) predicted longer PFS and OS in advanced NSCLC patients receiving single-agent ICI. The discrepancy in our results may be attributed to the fact that the serological indicators included in this study were only the baseline data and the different study endpoints.

Medical images serve as valuable sources of biomarkers, offering a macroscopic view of the tissue of interest (27). Radiomics has the ability to capture intricate details that are undetectable to the naked eye, and its efficacy has been demonstrated in numerous studies (28, 29). Besides, there remain several studies using radiomics to predict NIT in NSCLC patients. Yang et al. (11) constructed PET-based, CT-based, and combined PET-CT radiomics models to predict pCR to NIT in NSCLC, which achieved AUCs of 0.728, 0.732, and 0.818. Notably, the conventional CT-based radiomics model of the current study also demonstrated commendable results, especially the post-radiomics model. This aligns with clinical intuition—post-treatment images directly reflect early treatment-induced tumor changes, which are critical for predicting long-term response. Considering that the changes in the composition of lung cancer after treatment are difficult to be observed by the naked eye, delta-radiomics features were included in this study. Among the three delta-radiomics models, the delta1 model provided an unbiased perspective that reflects absolute difference without any consideration of relativity, and exhibited superior performance. Similarly, Gong et al.’s findings (30) showed that the predictive capability of delta-radiomics model for endpoint surpassed that of baseline (pre-treatment) radiomics model.

From a biological standpoint, rad-scores, derived from radiomics data, could reflect underlying tumor characteristics, such as heterogeneity and vascularization, which has been shown to correlate with treatment response (31). The rad-score in our combined model is determined by both post-treatment conventional radiomics and delta1-radiomics features. These features capture important tumor characteristics before and after treatment, providing a comprehensive evaluation of treatment efficacy. We analyze the most impactful features contributing to the rad-score, focusing on their respective coefficients and how they relate to treatment outcomes. Notably, delta1 features were shown to contribute significantly to predicting treatment outcomes, capturing dynamic changes in tumor characteristics over time. Among the most important delta1 features, delta1-log-sigma-5-mm-3D_firstorder_Minimum (coefficient: 1.4878) represents the minimum intensity value within the region of interest (ROI) after applying a 5-mm logarithmic filter. A post-treatment reduction in this value suggests a decrease in the lowest density regions, potentially indicating necrosis or diminished tumor cellularity, reflective of an effective therapeutic response. Similarly, delta1-wavelet-LLH_firstorder_10Percentile (coefficient: 1.0558), which reflects the 10th percentile of intensity values in the ROI post-wavelet transformation, decreases as low-density regions expand within the tumor, further implying treatment efficacy. Another important feature, delta1-log-sigma-4-mm-3D_ngtdm_Strength (coefficient: 1.0401), measures the smoothness of intensity variations in the tumor, with a post-treatment increase suggesting greater tissue homogeneity, indicative of successful tissue remodeling. In contrast, delta1-wavelet-LHH_glcm_ClusterShade (coefficient: -1.7578), a measure of pixel pair complexity, decreases in post-treatment images, implying reduced textural complexity and cellular heterogeneity, both of which were associated with positive therapeutic outcomes. (more details see in Supplementary Discussion) Overall, these radiomics features provide key insights into the tumor’s response to neoadjuvant immunotherapy, offering a quantitative approach for evaluating treatment efficacy and supporting clinical decision-making.

Deep learning has the capability to quantify high-throughput features that are beyond human perception. A previous study (32) suggested that the deep learning score may be associated with pathways involved in tumor proliferation and the enhancement of antitumor immune cell infiltration within the tumor microenvironment. For instance, Trebeschi et al. (33) utilized a deep learning imaging score to predict the prognosis of chemoimmunotherapy in advanced NSCLC, demonstrating that deep learning can capture both tumor- and non-tumor-related morphological changes. Similarly, She et al. (32) developed a deep learning model to predict major pathological response (MPR) in NSCLC patients treated with neoadjuvant ICIs combined with chemotherapy, revealing that higher deep learning scores were associated with an increasing likelihood of achieving MPR. Furthermore, Qu et al. (34) constructed a CT-based deep learning model to predict pCR to neoadjuvant immunotherapy in NSCLC, achieving AUCs of 0.775 in the validation set and 0.743 in the external cohort.

The AUCs of the deep learning model constructed in our study were 0.832 and 0.818 for training and internal testing cohorts, surpassing those of previously reported models mentioned above. As our experimental results showed, patients with high Deep-scores were more likely to achieve pCR. In the context of predicting treatment efficacy, Deep-scores have shown considerable potential in identifying key imaging patterns associated with therapeutic responses. In our predictive models, the inclusion of Deep-scores has significantly enhanced the model’s ability to forecast outcomes. Deep-scores are generated by processing high-dimensional imaging data through convolutional layers in neural networks, which may reflect biological processes like necrosis, fibrosis, or immune cell infiltration, all of which can occur in response to neoadjuvant immunotherapy that may not be easily discerned through traditional radiomics or clinical features (35). However, it is important to acknowledge that these Deep-scores should be interpreted as surrogate imaging patterns that indirectly reflect these biological phenomena rather than direct histopathological evidence. Unlike handcrafted radiomics features, Deep-scores leverage the model’s ability to automatically learn abstract and intricate representations from raw image data, making them especially valuable in capturing subtle variations in tumor structure and behavior over time. Its ability to capture the intricate interplay between different imaging characteristics allows for a more nuanced and accurate prediction of pCR. For instance, a decrease in heterogeneity or a shift in certain Deep-scores may indicate the tumor’s responsiveness to therapy, as regions of viable tumor tissue shrink or become more uniform in texture (36). Moreover, as Deep-scores aggregate multiple layers of imaging information, they may provide a more holistic view of tumor evolution, which can be crucial for distinguishing between responders and non-responders to treatment. Therefore, integrating Deep-scores into treatment efficacy models not only improves predictive performance but also provides a deeper understanding of how specific imaging features correlate with therapeutic success, offering valuable insights for clinicians in treatment planning.

While the results are promising, several limitations in the current study must be acknowledged. The primary limitation of this study is the relatively small sample size, particularly in the external validation cohort. While this may limit the representation of all clinical scenarios, the stable performance across different centers suggests the model’s robustness. We addressed potential overfitting through rigorous feature selection and transfer learning; however, future large-scale studies are necessary to confirm these findings. Additionally, as a retrospective analysis, selection bias is unavoidable, underscoring the need for prospective studies in the future. Moreover, this study focuses on short-term outcomes, and future studies should explore long-term endpoints to provide a more comprehensive understanding of the model’s effectiveness over time.

In conclusion, our results have demonstrated that the integration of radiomics features, deep learning features, and clinical information could offer superior predictive performance compared to using single-modality information alone. Specifically, the results have shown that the combination of these three sets of features could provide complementary insights, significantly improving the accuracy of pCR prediction in NSCLC patients undergoing NIT. These findings underscore the potential of multimodal integration in clinical predictive modeling and suggest that this approach could serve as the a point-of-care decision tool, enabling personalized treatment planning and ultimately improving clinical outcomes for NSCLC patients.