Immune checkpoint inhibitors (ICIs) have prolonged survival in patients with inoperable locally advanced non-small cell lung cancer (LA-NSCLC), and have become a standard therapy in current clinical practice guidelines (1, 2). However, clinical outcomes remain highly variable. While some patients experience long-lasting benefits, rapid disease progression is seen in others (3). This difference underscores the urgent need for better pre-treatment prediction of overall survival (OS) to optimize treatment decisions (4). By identifying high-risk patients early, physicians can potentially intensify treatment for those likely to derive little benefit from standard immunotherapy, or alternatively avoiding unnecessary overtreatment in low-risk patients (5).

To address this challenge, radiomics has emerged as a promising, non-invasive tool to analyze tumor phenotypes, predict OS, and guide treatments (6, 7). However, traditional radiomics typically extract hand-crafted features from the entire tumor volume, which implies we cannot spatially resolve phenotypical heterogeneity inside the tumor. A more effective approach turns out to be voxel-level radiomics, which enables us to derive heuristic spatial clusters from feature-by-feature “heatmaps”. While voxel level strategies have been explored in other cancer (8), their potential to uncover the complexity of LA-NSCLC and predict immunotherapy outcomes remains underexplored.

Moreover, implementing advanced deep learning models in the field of immunotherapy faces a challenge due to limited data availability (9). Deep learning models rely on large datasets to learn robust patterns, but immunotherapy has only become widely used in recent years, resulting in relatively small datasets (10). As a result, training complex models on these limited datasets increases the risk of overfitting and poor generalizability. To address this problem, we propose a novel knowledge transfer strategy (11). By leveraging extensive historical data from non-immunotherapy treatments, the Vision-Mamba can be first pre-trained in a supervised manner to learn stable tumor imaging representations, using overall survival as the training endpoint. This pre-training step can capture tumor characteristics that could be transferred to immunotherapy treated cohorts. Such pre-trained features can then be adapted to the specific context of immunotherapy, effectively bridging the data gap between small sample sizes and high-dimensional feature learning (12).

In this study, we developed a novel prognostic framework for inoperable LA-NSCLC by combining voxel-level deep radiomics with a cross-cohort training strategy. Specifically, by leveraging large, high-quality non-immunotherapy datasets for pre-training and subsequently fine-tuning on immunotherapy cohorts, our model was designed to provide accurate OS predictions with high generalizability.

In the study, we developed and validated a voxel-level deep radiomics framework to predict OS in patients with inoperable LA-NSCLC treated with immune checkpoint inhibitors. By leveraging tumor imaging representations learned from large-scale non-immunotherapy cohorts, and transferring these representations through selective fine-tuning of the Vision-Mamba architecture, the model effectively adapted knowledge encoded in deep imaging features to the immunotherapy setting. The stable discrimination and acceptable calibration observed in the independent immunotherapy testing set indicate that pretreatment CT imaging contains prognostic information that remains informative across different settings and supports the feasibility of applying imaging-based survival models in the immunotherapy era.

Our findings are consistent with prior radiomics and deep learning-based studies in NSCLC, which have reported associations between pretreatment imaging features and OS, including in immunotherapy treated populations (7, 13–16). However, most existing radiomics studies extract features at the whole tumor level, averaging spatial information across the tumor, or were developed and evaluated within relatively small immunotherapy dataset (10, 17). While these approaches have shown encouraging performance, their ability to generalize across evolving clinical settings remains uncertain (18). In contrast, our study focused on preserving intratumoral spatial information and explicitly evaluates generalization in an independent immunotherapy cohort, extending existing work toward more clinically heterogeneous populations. Existing clinical risk scores for LA-NSCLC, predominantly based on factors such as AJCC stage and performance status, provide a foundational prognostic framework. Our voxel-level deep radiomics model adds a complementary, imaging-derived dimension that quantifies intratumoral spatial heterogeneity. This may offer improved risk discrimination and pave the way for more integrated prognostic tools in the immunotherapy era.

The differences observed between the non-immunotherapy and immunotherapy datasets reflect the evolving clinical landscape of inoperable LA-NSCLC (Table 1; Supplementary Table 1). Changes in radiotherapy techniques and target volume are consistent with shifts in treatment strategies and patient selection accompanying the adoption of immunotherapy and therefore represent inherent heterogeneity in contemporary clinical practice (19, 20). Such heterogeneity poses a challenge for prediction of overall survival models developed within a single treatment context, as treatment patterns may limit generalizability. The preserved performance of our model in the immunotherapy testing dataset suggests that learning generalized imaging representations from historical data, followed by adaptation to patients treated with immunotherapy, is a reasonable method to address this issue. These findings underscore the need for modeling strategies that explicitly account for temporal and therapeutic heterogeneity when developing imaging based prognostic tools for modern oncology practice (21, 22).

The value of voxel-level image representation is highlighted by the model’s ability to capture spatially heterogeneous activation patterns within tumors, as visualized by voxel-level feature maps. Unlike global descriptors, voxel-level features preserve local spatial patterns (Supplementary Figure 2). Biologically, high intratumoral heterogeneity is increasingly recognized as a key driver of immune resistance and poor survival (23–25). By decoding these complex spatial phenotypes, our model leveraged spatial complexity for prognostic estimation. This implicit representation may contribute to coherent risk estimation across diverse patients.

Despite the high dimensionality inherent to voxel level radiomics and the limited size of the immunotherapy dataset, the final model achieved a C-index of 0.73 (Table 2), alongside acceptable calibration across time points (Figure 3). Such stability would be difficult to achieve if voxel level models were trained exclusively on small immunotherapy dataset, where overfitting and unstable risk estimation are common (22). These findings indicate that learning generalized imaging representations from large non-immunotherapy datasets before adaptation is not merely optional but necessary to fully leverage voxel level information in data-limited immunotherapy settings (26).

Regarding the feature selection process, the threshold of AUC > 0.6 was chosen empirically as a baseline filter to retain features with modest univariate prognostic signal while maintaining a manageable subset for downstream modeling. This selection was performed strictly within the training cohort using five fold cross-validation to avoid data leakage. We acknowledge that alternative thresholds or multivariate selection methods could be explored in future work to further optimize the balance between feature informativeness and model complexity.

The gradual decline in time-dependent AUC observed in the immunotherapy testing dataset with increasing time (Figure 2) is also clinically plausible. As follow-up extends, survival outcomes are increasingly influenced not only by post-treatment factors that are not captured by baseline imaging, such as subsequent lines of therapy and comorbid conditions, but also by potential biases related to outcome ascertainment during long-term follow-up (27, 28). In addition, the number of patients at risk decreases over time due to censoring and events, which can lead to greater variability in performance estimates. These factors suggest that baseline imaging-based models are inherently better suited for early survival prediction, and that declining long term discrimination reflects the growing impact of dynamic clinical variables, such as subsequent lines of therapy, rather than a limitation of the modeling method itself (29).

From a clinical perspective, the proposed model provides a practical tool for pre-treatment risk stratification in patients with inoperable LA-NSCLC receiving immunotherapy. The ability to identify patients at higher risk of poor survival before treatment initiation may support more individualized treatment planning, such as consideration of intensified multimodal strategies or closer post treatment surveillance (30), while sparing low risk patients from unnecessary escalation (28, 31). For example, a patient classified as high-risk by our model could be prioritized for enrollment in clinical trials exploring intensified immunotherapy regimens (e.g., combination with anti-angiogenic agents or novel immune modulators). Alternatively, such patients might benefit from intensified surveillance protocols, including more frequent follow-up imaging (e.g., CT scans every 2–3 months) to enable early detection of disease progression and prompt intervention. These actionable strategies exemplify how our model could translate into tailored clinical management. To facilitate the practical use of the model in multidisciplinary discussions, its continuous risk score could be translated into categorical tiers (e.g., Low/Intermediate/High) based on clinically meaningful survival probability thresholds, and visualized alongside the CT scans using color-coded scales and attention heatmaps. These actionable strategies exemplify how our model could translate into tailored clinical management. Future efforts should focus on prospective validation, integration with clinical imaging systems, and realworld evaluation of the model’s impact on clinical decision making and patient outcomes.

Several limitations of the study should be acknowledged. First, this study is retrospective in nature. We plan to conduct prospective trials to evaluate the model’s performance in real-world clinical settings, which will be essential for establishing its clinical utility and guiding potential implementation. Second, the use of 2D axial slice optimizes computational efficiency but discards 3D information. Third, variability in CT scanning parameters (e.g., slice thickness, reconstruction kernel) across institutions could affect the stability of radiomic features and the model’s generalizability in real-world practice. Fourth, patient comorbidities, variations in supportive care, and subsequent lines of therapy after progression, were not fully captured in our datasets and may act as unmeasured confounders affecting the model’s predictions and generalizability. Integrating temporal dynamic image features, and dosiomics may further enhance the predictive performance. Finally, key predictive biomarkers relevant to immunotherapy response, such as programmed death-ligand 1 (PD−L1) expression and tumor mutational burden (TMB), were not available in our retrospective cohorts and therefore could not be integrated into the model. Future studies that incorporate such molecular data alongside deep radiomic features hold promise for building more comprehensive and biologically informed prognostic tools. Although voxel-level features improved prognostic performance, the direct biological correlates of these spatially resolved imaging patterns remain incompletely understood. Future studies that integrate voxel-level radiomics with spatially mapped genomic or histopathological data are needed to establish clearer links between these imaging phenotypes and underlying tumor biology, which would enhance the model’s biological plausibility and clinical acceptability.

In conclusion, we present a robust prognostic framework for inoperable LA-NSCLC that combined voxel level radiomics with transfer learning. By effectively leveraging historical data to characterize intratumoral heterogeneity, the model provides a generalized, non-invasive tool for predicting overall survival in the immunotherapy era, paving the way for more personalized therapy. Future work should focus on prospective multi−center validation, integration with complementary biomarkers, and addressing practical implementation barriers to facilitate clinical translation.