CT-based deep learning for survival stratification in EGFR-mutant lung adenocarcinoma after EGFR-TKI resistance: A multicenter study Authors: Shuai Qie, Yasong Shi, Jingyun Li, Sicong Jia, Ming Liu Published: 2026 ABSTRACT SummaryAccurate risk assessment after EGFR-TKI resistance is important for guiding subsequent management of patients with EGFR-mutant lung adenocarcinoma. In this multicenter retrospective study, we developed and externally validated a computed tomography (CT)-based deep learning model using pretreatment CT images from 525 patients. A 2.5D ResNet-101 model showed consistent performance across training and external validation cohorts and enabled risk stratification for progression-free and overall survival. The deep learning score remained an independent prognostic factor after adjustment for clinical variables and demonstrated a continuous association with survival risk. These findings support the use of imaging-based deep learning approaches for individualized prognostic assessment in EGFR-mutant lung adenocarcinoma after EGFR-TKI resistance. Introduction EGFR-mutant lung adenocarcinoma is highly sensitive to EGFR tyrosine kinase inhibitors (TKIs); however, acquired resistance inevitably occurs.1 For patients who do not develop the T790M resistance mutation, subsequent treatment options remain limited, and clinical outcomes are suboptimal.2 In this context, immune checkpoint inhibitor (ICI) combined with chemotherapy has emerged as a potential treatment strategy, although its efficacy is highly heterogeneous and difficult to predict.3 Currently available biomarkers for predicting treatment outcomes in EGFR-mutant non-small cell lung cancer (NSCLC) are insufficient.4 PD-L1 expression has demonstrated inconsistent predictive value, and comprehensive molecular profiling is often constrained by limited tissue availability, cost, and feasibility in routine clinical practice.5 Therefore, there is a clear need for reliable, non-invasive tools to characterize tumor immune heterogeneity and guide treatment selection in this challenging population. Medical imaging offers a non-invasive means of capturing tumor characteristics in vivo.6 Recent advances in deep learning (DL) have enabled the identification of imaging patterns associated with tumor heterogeneity and treatment response.7,8 However, most prior studies have focused on treatment-naive populations or mutation prediction, and data remain limited for patients in the clinically relevant post-TKI resistance setting. Recent advances in DL have significantly improved prognostic modeling in lung cancer. In particular, convolutional neural network (CNN)-based frameworks and hybrid architectures have demonstrated strong performance in survival prediction tasks across multicenter cohorts. More recently, transformer-based models and hybrid CNN-transformer architectures, such as Vision Transformers (ViTs), ConvNeXt, and EfficientNet variants, have been explored to enhance feature representation and long-range dependency modeling. Despite these advances, many of these approaches require large-scale datasets and remain challenging to generalize in small-to-moderate clinical cohorts. Therefore, classical CNN architectures such as ResNet remain widely used due to their stability and robustness in medical imaging applications. Recent advances in DL have demonstrated promising performance in lung cancer prognosis. However, most existing studies are limited by single-center datasets or rely on a single model architecture, and few have systematically compared 2D, 2.5D, and 3D CNN architectures in the post-EGFR-TKI resistance setting with external multicenter validation. In this study, we developed and externally validated a computed tomography (CT)-based DL framework for survival stratification in patients with EGFR-mutant lung adenocarcinoma after resistance to first-generation EGFR-TKIs. We further investigated whether imaging-derived DL signatures could provide continuous prognostic assessment beyond conventional clinical variables. Results Clinical characteristics A total of 525 patients were included in this study, comprising 247 patients in the training cohort, 216 in external validation cohort 1, and 62 in external validation cohort 2. The median age of the overall population was 62 years (interquartile range [IQR], 55–68 years), and 62.9% of the patients were female. Most patients were never-smokers (74.3%) and had an Eastern Cooperative Oncology Group (ECOG) performance status of 0–1 (88.2%). The majority of patients were diagnosed at stage IV (85.9%), while 14.1% had stage III disease. Comorbidities were common, with 13.7% of patients having diabetes, 36.8% hypertension, and 21.7% heart disease. Regarding molecular characteristics, 38.9% of patients harbored EGFR exon 19 deletion, and 61.1% had L858R mutations. At baseline, brain metastases were present in 34.7% of the patients, liver metastases in 9.5%, and bone metastases in 48.2%. Baseline characteristics were largely comparable across cohorts. No significant differences were observed in age, sex, ECOG performance status, primary tumor location, smoking status, comorbidities, EGFR mutation subtype, or metastatic patterns (all p > 0.05). However, a significant difference in disease stage distribution was observed (p = 0.002), with a higher proportion of stage III patients in the validation cohorts. Detailed baseline characteristics are presented in Table 1. Characteristics stratified by DL-derived risk groups are summarized in Table S2, with no consistent pattern of imbalance observed across clinically relevant variables, suggesting limited confounding from baseline factors.Table 1Baseline clinicopathological characteristics of patients in the training and external validation cohortsAll (N = 525)Training cohort (N = 247)Validation cohort 1 (N = 216)Validation cohort 2 (N = 62)p valueSex−−−−0.122 Female330 (62.86%)162 (65.59%)125 (57.87%)43 (69.35%)− Male195 (37.14%)85 (34.41%)91 (42.13%)19 (30.65%)−Age62 (55, 68)62.00 (55, 69)62 (55, 68)63 (55, 68)0.88ECOG−−−−0.102 0–1463 (88.19%)210 (85.02%)196 (90.74%)57 (91.94%)− ≥ 262 (11.81%)37 (14.98%)20 (9.26%)5 (8.06%)−Stage−−−−0.002 III74 (14.10%)21 (8.50%)39 (18.06%)14 (22.58%)− IV451 (85.90%)226 (91.50%)177 (81.94%)48 (77.42%)−Primary site−−−−0.465 Left226 (43.05%)102 (41.30%)93 (43.06%)31 (50.00%)− Right299 (56.95%)145 (58.70%)123 (56.94%)31 (50.00%)−Diabetes−−−−0.457 No453 (86.29%)218 (88.26%)183 (84.72%)52 (83.87%)− Yes72 (13.71%)29 (11.74%)33 (15.28%)10 (16.13%)−Hypertension−−−−0.211 No332 (63.24%)161 (65.18%)138 (63.89%)33 (53.23%)− Yes193 (36.76%)86 (34.82%)78 (36.11%)29 (46.77%)−Heart disease−−−−0.089 No411 (78.29%)199 (80.57%)170 (78.70%)42 (67.74%)− Yes114 (21.71%)48 (19.43%)46 (21.30%)20 (32.26%)−Smoking−−−−0.664 No390 (74.29%)187 (75.71%)156 (72.22%)47 (75.81%)− Yes135 (25.71%)60 (24.29%)60 (27.78%)15 (24.19%)−BMI−−−−0.609. < 18.519 (3.62%)13 (5.26%)5 (2.31%)1 (1.61%)− 18.5–23.9299 (56.95%)137 (55.47%)124 (57.41%)38 (61.29%)− 24–27.9152 (28.95%)70 (28.34%)66 (30.56%)16 (25.81%)− > 2855 (10.48%)27 (10.93%)21 (9.72%)7 (11.29%)−EGFR subtype−−−−0.846 19 del204 (38.86%)93 (37.65%)87 (40.28%)24 (38.71%)− 21 L858R321 (61.14%)154 (62.35%)129 (59.72%)38 (61.29%)−Brain metastasis−−−−0.658 No343 (65.33%)163 (65.99%)137 (63.43%)43 (69.35%)− Yes182 (34.67%)84 (34.01%)79 (36.57%)19 (30.65%)−Liver metastasis−−−−0.749 No475 (90.48%)225 (91.09%)193 (89.35%)57 (91.94%)− Yes50 (9.52%)22 (8.91%)23 (10.65%)5 (8.06%)−Bone metastasis−−−−0.498 No272 (51.81%)123 (49.80%)113 (52.31%)36 (58.06%)− Yes253 (48.19%)124 (50.20%)103 (47.69%)26 (41.94%)−Abbreviations: ECOG, Eastern Cooperative Oncology Group; BMI, body mass index; EGFR, epidermal growth factor receptor. Table 1Baseline clinicopathological characteristics of patients in the training and external validation cohortsAll (N = 525)Training cohort (N = 247)Validation cohort 1 (N = 216)Validation cohort 2 (N = 62)p valueSex−−−−0.122 Female330 (62.86%)162 (65.59%)125 (57.87%)43 (69.35%)− Male195 (37.14%)85 (34.41%)91 (42.13%)19 (30.65%)−Age62 (55, 68)62.00 (55, 69)62 (55, 68)63 (55, 68)0.88ECOG−−−−0.102 0–1463 (88.19%)210 (85.02%)196 (90.74%)57 (91.94%)− ≥ 262 (11.81%)37 (14.98%)20 (9.26%)5 (8.06%)−Stage−−−−0.002 III74 (14.10%)21 (8.50%)39 (18.06%)14 (22.58%)− IV451 (85.90%)226 (91.50%)177 (81.94%)48 (77.42%)−Primary site−−−−0.465 Left226 (43.05%)102 (41.30%)93 (43.06%)31 (50.00%)− Right299 (56.95%)145 (58.70%)123 (56.94%)31 (50.00%)−Diabetes−−−−0.457 No453 (86.29%)218 (88.26%)183 (84.72%)52 (83.87%)− Yes72 (13.71%)29 (11.74%)33 (15.28%)10 (16.13%)−Hypertension−−−−0.211 No332 (63.24%)161 (65.18%)138 (63.89%)33 (53.23%)− Yes193 (36.76%)86 (34.82%)78 (36.11%)29 (46.77%)−Heart disease−−−−0.089 No411 (78.29%)199 (80.57%)170 (78.70%)42 (67.74%)− Yes114 (21.71%)48 (19.43%)46 (21.30%)20 (32.26%)−Smoking−−−−0.664 No390 (74.29%)187 (75.71%)156 (72.22%)47 (75.81%)− Yes135 (25.71%)60 (24.29%)60 (27.78%)15 (24.19%)−BMI−−−−0.609. < 18.519 (3.62%)13 (5.26%)5 (2.31%)1 (1.61%)− 18.5–23.9299 (56.95%)137 (55.47%)124 (57.41%)38 (61.29%)− 24–27.9152 (28.95%)70 (28.34%)66 (30.56%)16 (25.81%)− > 2855 (10.48%)27 (10.93%)21 (9.72%)7 (11.29%)−EGFR subtype−−−−0.846 19 del204 (38.86%)93 (37.65%)87 (40.28%)24 (38.71%)− 21 L858R321 (61.14%)154 (62.35%)129 (59.72%)38 (61.29%)−Brain metastasis−−−−0.658 No343 (65.33%)163 (65.99%)137 (63.43%)43 (69.35%)− Yes182 (34.67%)84 (34.01%)79 (36.57%)19 (30.65%)−Liver metastasis−−−−0.749 No475 (90.48%)225 (91.09%)193 (89.35%)57 (91.94%)− Yes50 (9.52%)22 (8.91%)23 (10.65%)5 (8.06%)−Bone metastasis−−−−0.498 No272 (51.81%)123 (49.80%)113 (52.31%)36 (58.06%)− Yes253 (48.19%)124 (50.20%)103 (47.69%)26 (41.94%)−Abbreviations: ECOG, Eastern Cooperative Oncology Group; BMI, body mass index; EGFR, epidermal growth factor receptor. Baseline clinicopathological characteristics of patients in the training and external validation cohorts Abbreviations: ECOG, Eastern Cooperative Oncology Group; BMI, body mass index; EGFR, epidermal growth factor receptor. Model performance The performance of multiple DL configurations, including 2D, 2.5D, and 3D architectures based on different imaging views, was systematically evaluated. Among these, the 2.5D axial model demonstrated the best and most stable discriminative performance across all cohorts. Specifically, the 2.5D axial model achieved area under the curve (AUC) values of 0.907 (95% confidence interval [CI]: 0.860–0.955), 0.904 (95% CI: 0.831–0.976), and 0.863 (95% CI: 0.780–0.946) in the training cohort, external validation cohort 1, and external validation cohort 2, respectively (Figure 1A).Figure 1Performance of deep learning models across cohorts(A) ROC curves of the 2D, 2.5D, and 3D models in the training cohort, external validation cohort 1, and external validation cohort 2.(B) DCA demonstrating the net clinical benefit of different models across a range of threshold probabilities.(C) Calibration curves showing agreement between predicted and observed outcomes across cohorts. Performance of deep learning models across cohorts (A) ROC curves of the 2D, 2.5D, and 3D models in the training cohort, external validation cohort 1, and external validation cohort 2. (B) DCA demonstrating the net clinical benefit of different models across a range of threshold probabilities. (C) Calibration curves showing agreement between predicted and observed outcomes across cohorts. The optimal cutoff value for the DL score was determined using the Youden index in the training cohort and was subsequently applied to stratify patients in all cohorts. Pairwise comparisons using the DeLong test demonstrated that the 2.5D axial model consistently outperformed the 2D models across all cohorts. In the training cohort, the AUC difference between the 2.5D axial and 2D axial models was 0.084 (p = 0.003). Similar improvements were observed in external validation cohort 1 (AUC difference: 0.091, p = 0.012) and external validation cohort 2 (AUC difference: 0.073, p = 0.028). Although the differences between the 2.5D and 3D models were not consistently significant, the 2.5D model demonstrated more stable generalization performance and calibration across independent cohorts (Table S3; Figure S1). In contrast, the performance of the 3D model was less consistent and showed reduced discriminative ability in external validation cohort 1. No statistically significant difference was observed between the 2.5D and 3D models in external validation cohort 2; however, the 2.5D model demonstrated superior overall robustness and calibration consistency across the multicenter datasets. Overall, the performance differences between the 2D and 3D models were modest and inconsistent across cohorts (Table S4). Decision curve analysis (DCA) demonstrated that the 2.5D model provided a higher net benefit than the 2D and 3D models across a wide range of threshold probabilities in all cohorts, indicating superior clinical utility (Figure 1B). Calibration curves showed good agreement between predicted and observed outcomes for the 2.5D model across all cohorts, whereas the 2D model exhibited relatively poorer calibration performance (Figure 1C). Collectively, these findings support the selection of the 2.5D axial architecture as the final DL model, given its superior discriminative performance, improved calibration, and more stable generalization capability across independent multicenter cohorts. Survival analysis Kaplan-Meier survival analysis demonstrated that patients in the high-risk group exhibited significantly worse progression-free survival (PFS) and overall survival (OS) compared with those in the low-risk group across the training and external validation cohorts (all log rank p < 0.001; Figures 2A–2C). Patients were stratified into high- and low-risk groups using the optimal cutoff value (0.501) determined by the Youden index in the training cohort. In the training cohort, the median PFS was significantly shorter in the high-risk group than in the low-risk group (5.78 months vs. 12.62 months; hazard ratio [HR], 2.163; 95% CI, 1.254–3.730; p = 0.006). Similarly, the high-risk group exhibited inferior OS compared with the low-risk group (16.8 months vs. 21.8 months; HR, 2.213; 95% CI, 1.270–3.855; p = 0.005). Consistent results were observed in the validation cohorts. In validation cohort 1, the median PFS was 7.69 months in the high-risk group and 17.76 months in the low-risk group (HR, 2.341; 95% CI, 1.289–4.250; p = 0.005), while the median OS was 15.6 months and 18.4 months, respectively (HR, 2.345; 95% CI, 1.299–4.233; p = 0.005). In external validation cohort 2, the median PFS was 7.92 months in the high-risk group and 13.44 months in the low-risk group (HR, 2.543; 95% CI, 1.298–4.985; p = 0.007), and the median OS was 13.9 months versus 19.3 months (HR, 1.890; 95% CI, 1.045–3.418; p = 0.035). To further characterize the relationship between the DL score and survival risk beyond dichotomous stratification, restricted cubic spline (RCS) analysis was performed. The DL score was modeled as a continuous variable, with the reference value set at the median DL score. For PFS, a significant nonlinear association was observed between the DL score and the risk of progression (p for overall < 0.001; p for nonlinearity < 0.001), with the HR increasing progressively at higher DL score levels (Figure 3A). For OS, a similar nonlinear relationship was identified, with increasing DL scores associated with a higher risk of mortality (p for overall <0.001; p for nonlinearity < 0.001) (Figure 3B). These findings indicate that the DL score captures a continuous and nonlinear relationship with survival risk.Figure 2Risk stratification and survival analysis based on the deep learning scoreKaplan-Meier survival curves for progression-free survival and overall survival stratified by the deep learning-derived risk score in the (A) training cohort, (B) external validation cohort 1, and (C) external validation cohort 2. Hazard ratios (HRs) with 95% CIs and log rank p values are shown for each cohort.Figure 3Continuous association between the DL score and survival outcomes using restricted cubic spline analysis(A) Relationship between the DL score and PFS.(B) Relationship between the DL score and OS. Solid lines represent HRs, and shaded areas indicate 95% CI. Solid lines represent HRs, and shaded areas indicate 95% CI. P values for overall association and nonlinearity were derived from the restricted cubic spline analysis. Risk stratification and survival analysis based on the deep learning score Kaplan-Meier survival curves for progression-free survival and overall survival stratified by the deep learning-derived risk score in the (A) training cohort, (B) external validation cohort 1, and (C) external validation cohort 2. Hazard ratios (HRs) with 95% CIs and log rank p values are shown for each cohort. Continuous association between the DL score and survival outcomes using restricted cubic spline analysis (A) Relationship between the DL score and PFS. (B) Relationship between the DL score and OS. Solid lines represent HRs, and shaded areas indicate 95% CI. Solid lines represent HRs, and shaded areas indicate 95% CI. P values for overall association and nonlinearity were derived from the restricted cubic spline analysis. Although the model achieved a high sensitivity for identifying high-risk patients, a trade-off with specificity was observed, which may contribute to potential false-positive classifications. This behavior is consistent with the model’s emphasis on maximizing sensitivity for clinical risk stratification. Independent prognostic analysis To evaluate whether the DL-derived risk score was an independent predictor of survival outcomes, univariate and multivariate Cox proportional hazards regression analyses were performed in the training cohort and independent external validation cohorts. For PFS, in the training cohort, the DL-based risk stratification was significantly associated with PFS in univariate analysis (HR, 1.779; 95% CI, 1.060–2.985; p = 0.029) and remained an independent predictor after multivariate adjustment (HR, 2.163; 95% CI, 1.254–3.730; p = 0.006). Consistent results were observed in the external validation cohorts. In validation cohort 1, the DL score remained significantly associated with PFS in multivariate analysis (HR, 2.341; 95% CI, 1.289–4.250; p = 0.005). Similarly, in validation cohort 2, the DL score demonstrated consistent prognostic value (HR, 2.543; 95% CI, 1.298–4.985; p = 0.007). For OS, the DL-derived risk score also showed consistent prognostic significance across cohorts. In the training cohort, the DL score remained an independent predictor in multivariate analysis (HR, 2.213; 95% CI, 1.270–3.855; p = 0.005). This finding was validated in external cohort 1 (HR, 2.345; 95% CI, 1.299–4.233; p = 0.005) and external cohort 2 (HR, 1.890; 95% CI, 1.045–3.418; p = 0.035). Across all cohorts, the DL-derived risk score remained an independent prognostic factor after adjustment for the clinical covariates retained in the respective multivariable Cox regression models. While certain clinical factors, such as bone metastasis and smoking status, showed cohort-dependent associations with PFS, their effects were less consistent and did not remain significant across all datasets after multivariable adjustment. Because variable selection was performed independently for each cohort based on the results of the univariate analyses, the covariates retained in the final multivariable Cox regression models differed across cohorts (Tables S5–S10). Nevertheless, the DL-derived risk score remained statistically significant across all multivariable analyses, supporting its robust prognostic value across different clinical settings. Detailed results of the Cox regression analyses are provided in Tables S5–S10. Model interpretability To further explore the interpretability of the DL model, gradient-weighted class activation mapping (Grad-CAM) visualization was performed, and representative examples are shown in Figure 4. The activation maps indicated that the model focused on both intratumoral regions and adjacent peritumoral areas. These findings suggest that imaging features derived from both tumor characteristics and surrounding tissue context contributed to the DL-based risk prediction, providing qualitative support for the biological plausibility of the model. The spatial correspondence between high-attention regions and tumor areas further supports the reliability of the model. Representative cases further demonstrated that the predicted class was consistent with the ground-truth label, with high prediction confidence in correctly classified cases. The Grad-CAM heatmaps indicated that the model primarily focused on intratumoral and adjacent peritumoral regions. These attention regions spatially corresponded to the tumor boundary and surrounding tissue, suggesting that both the tumor morphology and the local microenvironment contributed to model predictions. Importantly, these qualitative findings are consistent with the quantitative survival analyses, in which the DL-derived risk score effectively stratified patients into distinct high- and low-risk groups with significantly different PFS and OS outcomes across all cohorts.Figure 4Model interpretability using Grad-CAM visualizationRepresentative Grad-CAM visualizations illustrating model interpretability. For each representative case, the original CT image, ground-truth label, predicted class, prediction confidence score, Grad-CAM heatmap, and overlay are shown. Warmer colors indicate regions contributing more strongly to model predictions. The highlighted regions were predominantly located within the tumor and adjacent peritumoral tissues, supporting the biological plausibility of the proposed framework. Model interpretability using Grad-CAM visualization Representative Grad-CAM visualizations illustrating model interpretability. For each representative case, the original CT image, ground-truth label, predicted class, prediction confidence score, Grad-CAM heatmap, and overlay are shown. Warmer colors indicate regions contributing more strongly to model predictions. The highlighted regions were predominantly located within the tumor and adjacent peritumoral tissues, supporting the biological plausibility of the proposed framework. Discussion This multicenter study demonstrates that a CT-based DL score provides consistent prognostic stratification in patients with EGFR-mutant lung adenocarcinoma after resistance to first-generation EGFR-TKIs. The model, particularly the 2.5D axial architecture, showed consistent discriminative performance across independent external cohorts and was associated with both PFS and OS. Importantly, the DL score retained independent prognostic value after adjustment for conventional clinical variables, supporting its potential clinical utility in this setting. Although more recent architectures, such as ViTs and lightweight CNN variants, have demonstrated promising performance in large-scale datasets, their effectiveness in relatively small-to-moderate medical imaging cohorts remains limited due to data-hungry characteristics and training instability. In contrast, ResNet-based architectures remain widely adopted in clinical imaging studies due to their robustness and reproducibility, making them suitable for fair architectural comparisons in this work. A key strength of this work is the systematic comparison of model architectures. Among the 2D, 2.5D, and 3D configurations across multiple views, the 2.5D axial model achieved the most stable performance across cohorts. This finding is biologically and methodologically plausible. Pure 2D models may fail to capture inter-slice contextual information, whereas fully 3D models are often limited by higher parameter complexity and sensitivity to data heterogeneity.9,10,11 The 2.5D approach offers a practical balance, preserving spatial context while maintaining robustness and generalizability.12,13 The consistent performance observed across external cohorts further supports the suitability of this architecture for real-world applications. Compared with recent state-of-the-art approaches for lung cancer imaging analysis, the proposed framework has several distinctive features. Recent studies have explored radiomics, deep radiomic features, semi-supervised learning, and transformer-based architectures for lung cancer prognosis and treatment response prediction. These approaches have improved feature representation and model flexibility, particularly when large-scale datasets or multimodal information are available. However, many existing models remain limited by single-center designs, heterogeneous endpoints, or insufficient external validation. In contrast, our study focused on a clinically specific and challenging population of EGFR-mutant lung adenocarcinoma after EGFR-TKI resistance and externally validated the model in two independent cohorts. Rather than emphasizing a newly designed backbone architecture, our contribution lies in systematically comparing 2D, 2.5D, and 3D CT-based representations under a unified ResNet-101 framework, demonstrating that the 2.5D axial model provides the most stable balance between spatial context, model complexity, and multicenter generalizability. The proposed model also differs from conventional radiomics-based approaches by learning hierarchical imaging representations directly from CT images, thereby reducing dependence on predefined handcrafted features. Compared with transformer-based models, which may require larger training datasets and more extensive computational resources, the ResNet-based 2.5D strategy provides a practical and reproducible solution for moderate-sized clinical cohorts. Nevertheless, newer architectures such as ConvNeXt, EfficientNet, and ViTs may further improve feature representation in larger datasets, and their integration with clinical, pathological, and genomic information should be investigated in future studies. Beyond dichotomous risk stratification, the present study also explored the continuous prognostic effect of the DL score. RCS analysis revealed a significant nonlinear association between the DL score and survival risk, indicating that the prognostic impact of imaging-derived features may not be adequately captured by a single cutoff. Notably, risk increased progressively at higher DL score levels, suggesting that the DL score encodes gradations of tumor aggressiveness rather than a simple binary state. These findings highlight the added value of modeling imaging biomarkers as continuous variables and support the complementary role of DL-based risk estimation alongside traditional categorical approaches. The interpretability analysis provides additional insight into the imaging features captured by the DL model. Grad-CAM visualizations suggested that the model attended to both intratumoral and peritumoral regions, implying that information related to tumor heterogeneity and the surrounding microenvironment may contribute to risk prediction. Given that EGFR-TKI resistance is associated with dynamic changes in tumor biology and microenvironmental interactions,14,15 these imaging patterns may partially reflect underlying biological processes. However, further studies integrating radiogenomic or pathological data are needed to validate these observations. From a biological perspective, the observed prognostic associations likely reflect underlying tumor heterogeneity and microenvironmental alterations following EGFR-TKI resistance.16 Prior studies have suggested that resistance to EGFR-targeted therapy is accompanied by dynamic changes in tumor cellularity, vascularity, and immune infiltration.17,18 However, these interpretations remain speculative and require validation through radiogenomic or pathological correlation. Such alterations may manifest as subtle but quantifiable patterns on CT imaging.19,20 DL models, by learning hierarchical representations directly from imaging data, may capture these complex phenotypic signatures more effectively than handcrafted features. This may explain the independent prognostic value of the DL score observed in the current study. Clinically, accurate risk stratification in this population remains challenging, particularly in patients without the T790M mutation who are candidates for subsequent systemic therapies. The proposed DL model provides a non-invasive and readily accessible tool that could assist in identifying high-risk patients who may benefit from closer monitoring or alternative therapeutic strategies. In addition, the ability of the DL score to provide continuous risk estimates may facilitate more individualized decision-making. The relatively high sensitivity (∼85%) observed in this study may be associated with an increased false-positive rate, particularly in the context of a limited external validation cohort. From a clinical perspective, this trade-off reflects the model’s tendency to prioritize the identification of high-risk patients, which is often desirable in prognostic stratification tasks for oncology. To mitigate potential false-positive predictions and improve robustness, several strategies can be considered. First, threshold optimization based on the Youden index may be further refined according to specific clinical scenarios. Second, calibration techniques, such as probability calibration or decision threshold adjustment, can improve the alignment between predicted risk scores and actual outcomes. Third, future studies incorporating larger and more diverse multicenter cohorts may help reduce dataset-induced variability and improve generalization performance. Finally, integration of multimodal clinical variables may further enhance specificity while maintaining high sensitivity. Several limitations should be acknowledged. First, the retrospective design may introduce selection bias despite the use of multicenter data. Second, although external validation was performed, all cohorts were derived from institutions within a similar healthcare system, and further validation in more diverse populations is warranted. Third, biological interpretation of DL features remains limited, and future studies incorporating radiogenomic or pathological correlation may improve interpretability. Finally, although the model demonstrated prognostic value, its predictive value for specific treatment responses was not directly evaluated and requires further investigation. Additionally, the relatively small sample size in external validation cohort 2 may limit the stability of effect estimates. It should be noted that this study shares a partially overlapping patient cohort with our previous work; however, the present analysis addresses a fundamentally different clinical and methodological question. Specifically, the current study focuses on DL-based survival stratification using imaging-derived representations, whereas our prior work was designed to predict treatment response using radiomics-based and habitat imaging approaches. Importantly, the analytical frameworks, model architectures, endpoints, and outputs differ substantially between these studies. The present model generates a continuous risk score for survival outcomes (PFS and OS), capturing nonlinear prognostic associations, while previous models were primarily developed for classification tasks related to treatment response. In addition, no overlapping analyses, figures, or reported results are shared between the studies. Therefore, although the studies are derived from a similar clinical population, they provide complementary rather than redundant insights, reflecting distinct perspectives on tumor characterization and outcome prediction. This approach is consistent with emerging research paradigms in medical imaging, where multiple analytical strategies are applied to the same cohort to address different clinically relevant questions. In conclusion, this study demonstrates that a CT-based DL model enables robust prognostic stratification in patients with EGFR-mutant lung adenocarcinoma after EGFR-TKI resistance. The consistent performance across external cohorts underscores its generalizability and potential clinical utility. These findings highlight the value of imaging-based DL biomarkers for individualized risk assessment and continuous prognostic evaluation in this challenging clinical setting. Limitations of the study Several limitations of this study should be acknowledged. Although the proposed ResNet-101-based model demonstrated strong prognostic performance, its relatively high architectural complexity may limit scalability and clinical deployment in resource-constrained settings, and future work may explore more lightweight or hybrid architectures. In addition, despite the use of multicenter data, potential heterogeneity across participating institutions, including differences in CT acquisition protocols, scanner types, and patient populations, may still affect model generalizability, although external validation in two independent cohorts partially mitigates this concern. Furthermore, while Grad-CAM was used to improve interpretability, the overall explainability of DL models remains limited, and more advanced explainable AI methods are needed to enhance clinical transparency. Future research should focus on multimodal integration combining imaging, clinical, pathological, and genomic data to improve predictive performance and biological interpretability, as well as on the application of transformer-based architectures and large-scale prospective multicenter validation to further enhance robustness and clinical applicability. Although the proposed model demonstrated robust prognostic value for PFS and OS, it was not designed to directly predict treatment response or to evaluate whether model-guided clinical decisions could improve patient outcomes. Therefore, its predictive utility for selecting specific therapeutic strategies remains unproven. Future studies should prospectively assess whether a DL-based risk stratification can guide treatment intensification, closer surveillance, or alternative systemic therapies in high-risk patients. In addition, prospective validation in larger and geographically diverse populations is required to confirm clinical applicability beyond retrospective cohorts derived from institutions within a similar healthcare system. Resource availability Lead contact Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Ming Liu (36500653@hebmu.edu.cn). Materials availability This study did not generate or use any new experimental materials, reagents, or biological resources that require distribution. Data and code availability •The clinical imaging datasets are not publicly available because they contain identifiable patient information and are subject to institutional ethics and privacy regulations. Access may be granted only after approval by the participating institutions and ethics committee.•The DL code is publicly accessible at: https://github.com/qieshuai1984/EGFR-DL-survival.•Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. •The clinical imaging datasets are not publicly available because they contain identifiable patient information and are subject to institutional ethics and privacy regulations. Access may be granted only after approval by the participating institutions and ethics committee. The clinical imaging datasets are not publicly available because they contain identifiable patient information and are subject to institutional ethics and privacy regulations. Access may be granted only after approval by the participating institutions and ethics committee. •The DL code is publicly accessible at: https://github.com/qieshuai1984/EGFR-DL-survival. The DL code is publicly accessible at: https://github.com/qieshuai1984/EGFR-DL-survival. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Acknowledgments This work was supported by the Hebei Key Laboratory of Precise Imaging of Inflammation-Related Tumors (grant no. 2363P034). The funder had no role in study design, data collection, data analysis, data interpretation, manuscript preparation, or the decision to submit the article. Author contributions S.Q., conceptualization, methodology, software, formal analysis, investigation, data curation, visualization, and writing – original draft; Y.S., investigation, data curation, and writing – review and editing; J.L., software, data curation and writing – review and editing; S.J., validation, data curation, and writing – review and editing; M.L., conceptualization, supervision, writing – review and editing, project administration, and funding acquisition. Declaration of interests The authors declare no competing interests. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERDeposited dataClinical imaging datasetsHebei Medical University Third Hospital and collaborating institutionsDe-identified clinical CT datasets; access restricted by institutional ethics and privacy regulationsDeep learning model codeGitHubhttps://github.com/qieshuai1984/EGFR-DL-survivalSoftware and algorithmsResNet-101 architectureHe et al., Deep Residual Learning for Image RecognitionResNet-101 backbone used for deep learning model developmentMedical image segmentation softwareITK-SNAPVersion 3.8.0Statistical analysis softwareIBM SPSS StatisticsVersion 20.0Programming environmentPython Software FoundationPython version 3.9.1 REAGENT or RESOURCESOURCEIDENTIFIERDeposited dataClinical imaging datasetsHebei Medical University Third Hospital and collaborating institutionsDe-identified clinical CT datasets; access restricted by institutional ethics and privacy regulationsDeep learning model codeGitHubhttps://github.com/qieshuai1984/EGFR-DL-survivalSoftware and algorithmsResNet-101 architectureHe et al., Deep Residual Learning for Image RecognitionResNet-101 backbone used for deep learning model developmentMedical image segmentation softwareITK-SNAPVersion 3.8.0Statistical analysis softwareIBM SPSS StatisticsVersion 20.0Programming environmentPython Software FoundationPython version 3.9.1 Experimental model and study participant details Study participants This multicenter retrospective study included 525 patients with advanced (stage IIIB–IV) EGFR-mutant lung adenocarcinoma who had developed resistance to first-generation EGFR tyrosine kinase inhibitors (TKIs) and subsequently received immune checkpoint inhibitor (ICI) plus chemotherapy. Patients were recruited from three institutions: 247 patients from Hebei University Affiliated Hospital (training cohort), 216 patients from Baoding First Central Hospital (external validation cohort 1), and 62 patients from Hebei Medical University Third Hospital (external validation cohort 2). All patients were consecutively screened between April 1, 2019, and March 31, 2022, with a minimum follow-up duration of 36 months by the final data cutoff (March 31, 2025). The inclusion criteria were as follows: (1) histologically confirmed advanced lung adenocarcinoma; (2) presence of a sensitizing EGFR mutation, including exon 19 deletion or L858R mutation; (3) radiologically confirmed disease progression after first-line treatment with first-generation EGFR-TKIs according to RECIST version 1.1; (4) initiation of ICI plus chemotherapy after TKI resistance; and (5) availability of thin-section chest CT images (slice thickness ≤1.5 mm) acquired within 30 days before ICI treatment initiation. The exclusion criteria were as follows: (1) presence of the EGFR T790M resistance mutation or other actionable driver mutations, including ALK, ROS1, and RET alterations; (2) CT images with significant artifacts or slice thickness >1.5 mm; (3) insufficient follow-up data for PFS assessment; (4) history of other active malignancies within the previous 5 years; and (5) prior thoracic radiotherapy that could interfere with target lesion evaluation. Patients with incomplete clinical or imaging data were excluded. External validation cohorts were derived from independent institutions, and no patient overlap existed among cohorts. One pretreatment CT examination was included for each patient, resulting in a final dataset of 525 CT examinations from 525 individual patients (Below Figure).Study flowchart of patient selection Study flowchart of patient selection This study was approved by the institutional ethics committee of Hebei Medical University Third Hospital (approval number: HDFY-LL-2022-075). The requirement for written informed consent was waived due to the retrospective nature of the study. Molecular profiling and treatment eligibility All patients underwent molecular profiling after progression on first-generation EGFR-TKI therapy. Patients were included only if there was no evidence of the EGFR T790M resistance mutation in their medical records. The absence of other actionable driver mutations, including ALK, ROS1, and RET alterations, was confirmed according to the predefined exclusion criteria. Molecular profiling was primarily performed using tissue-based next-generation sequencing (NGS) according to institutional protocols. When tumor tissue was unavailable or insufficient, circulating tumor DNA analysis (liquid biopsy) was used as an alternative. This selection strategy ensured that the study population represented patients with EGFR-mutant lung adenocarcinoma who had developed resistance to first-generation EGFR-TKIs without T790M mutation and were candidates for subsequent ICI plus chemotherapy treatment. Clinical and pathological data collection Comprehensive clinical and pathological data were retrospectively collected for all enrolled patients. The collected variables included age, sex, smoking history, Eastern Cooperative Oncology Group (ECOG) performance status, pretreatment clinical stage, tumor location, and EGFR mutation subtype (exon 19 deletion or L858R mutation). Tumor staging was performed according to the 8th edition of the TNM classification system established by the International Association for the Study of Lung Cancer (IASLC). Method details Pretreatment evaluation Prior to initiation of immune-checkpoint inhibitor combined with chemotherapy (ICI-chemotherapy), a comprehensive pretreatment evaluation was performed for all patients to establish baseline clinical characteristics and ensure treatment eligibility. ECOG performance status was recorded before treatment initiation. Radiographic tumor assessment was performed within 4 weeks before treatment commencement using contrast-enhanced CT of the chest, abdomen, and pelvis. Brain imaging with magnetic resonance imaging (MRI) or CT was performed as required. Laboratory examinations included complete blood count, comprehensive metabolic panel, and screening for hepatitis B and C serology. This standardized pretreatment evaluation was performed to ensure adequate organ function and to establish baseline information for subsequent treatment and outcome assessment. Treatment regimens Following confirmed disease progression after first-generation EGFR-TKI therapy, all enrolled patients received immune checkpoint inhibitor (ICI) combined with chemotherapy as subsequent systemic treatment in routine clinical practice. All patients underwent molecular testing after TKI resistance, and only patients without evidence of the EGFR T790M resistance mutation were included. These patients were therefore not eligible for subsequent third-generation EGFR-TKI therapy, such as osimertinib. Treatment regimens were administered according to prevailing clinical guidelines and physician discretion. The most commonly used regimen consisted of a platinum-based doublet chemotherapy (carboplatin or cisplatin) combined with pemetrexed and an immune checkpoint inhibitor (pembrolizumab or sintilimab). Treatment was delivered in 3-week cycles. Although minor variations in treatment regimens existed among patients, the overall therapeutic strategy was consistent across the study population, reflecting a real-world clinical setting. Outcome assessment and follow-up Tumor response was assessed according to the Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1 by board-certified radiologists who were blinded to clinical outcomes. To ensure consistency, imaging evaluations were independently performed and reviewed according to standardized protocols across participating centers, and discrepancies were resolved through consensus review. Radiological assessments were conducted every two treatment cycles (approximately every 6 weeks) during the first 6 months and every 3 months thereafter until disease progression or death. The primary endpoint of this study was progression-free survival (PFS), defined as the time from initiation of ICI plus chemotherapy to the first documented radiological disease progression or death from any cause, whichever occurred first. Patients without disease progression or death were censored at the date of the last radiographic assessment. Secondary endpoints included objective response rate (ORR) and overall survival (OS). ORR was defined as the proportion of patients achieving a best overall response of complete response (CR) or partial response (PR) during treatment. OS was defined as the time from initiation of ICI plus chemotherapy to death from any cause, with patients censored at the last follow-up if alive. CT acquisition and preprocessing Contrast-enhanced chest CT scans were obtained from three participating centers using institutional standard acquisition protocols. All CT examinations were performed during a single breath-hold at full inspiration, covering the thorax from the lung apices to the costophrenic angles. Detailed CT acquisition parameters are provided in . All CT images were retrieved in Digital Imaging and Communications in Medicine (DICOM) format and underwent preprocessing before deep learning model development. To reduce variability among scanners and institutions, CT images were resampled to a uniform voxel spacing using linear interpolation, and intensity values were normalized. Tumor regions of interest (ROIs) were manually segmented on axial CT images using ITK-SNAP (version 3.8.0). Segmentation was performed independently by two experienced physicians and subsequently reviewed by a senior radiologist. In cases of discrepancy, consensus was reached through joint review. To evaluate segmentation reproducibility, inter-observer agreement was assessed in a randomly selected subset of 50 patients using the intraclass correlation coefficient (ICC). An ICC value ≥ 0.75 was considered indicative of good agreement. Following segmentation, tumor regions were cropped and resized to a standardized input size for subsequent model development. Deep learning model development Overview of model construction A deep learning workflow including image preprocessing, model construction, training, and risk score generation was developed to predict progression-free survival (PFS) using pretreatment CT images (Below Figure).Overview of the deep learning workflow Overview of the deep learning workflow Three deep learning architectures based on the ResNet-101 backbone were developed and compared, including 2D, 2.5D, and 3D models. For the 2D model, tumor regions were extracted from axial, sagittal, and coronal planes. For the 2.5D model, multiple adjacent axial slices centered on the tumor were incorporated to provide additional contextual information while maintaining computational efficiency. For the 3D model, volumetric tumor regions were directly used as network inputs to preserve three-dimensional spatial information. All tumor regions of interest were cropped and resized to a standardized input size before model training. ResNet-101 backbone selection ResNet-101 was selected as the backbone network because of its established robustness and reproducibility in medical imaging applications, particularly for datasets of moderate size. Compared with more recent architectures, including MobileNet, EfficientNet, ConvNeXt, and Vision Transformers (ViTs), ResNet-101 provides a balanced trade-off between model complexity and generalization performance. The aim of this study was not to develop a novel backbone architecture, but rather to systematically compare different image representation strategies (2D, 2.5D, and 3D) under a unified and widely validated CNN framework. Dataset partitioning and data augmentation Dataset partitioning was performed at the patient level to prevent information leakage. All CT images from an individual patient were assigned exclusively to a single cohort (training or external validation), ensuring that no patient contributed images to more than one dataset. One pretreatment CT examination was included for each patient. Data augmentation was applied exclusively to the training cohort and included online random flipping and random rotation during model training. No augmentation was applied to the external validation cohorts. Model training and optimization The models were optimized using the Adam optimizer with an initial learning rate of 1 × 10−4 and a batch size of 32. The maximum number of training epochs was 30, and binary cross-entropy was used as the loss function. The ResNet-101 backbone was initialized with ImageNet-pretrained weights and subsequently fine-tuned using the CT dataset. ReLU activation functions were used throughout the convolutional layers, whereas a sigmoid activation function was applied to the output layer for risk estimation. To reduce overfitting, regularization strategies including dropout (rate = 0.5), data augmentation, and early stopping based on validation performance were adopted. Training loss and validation performance were continuously monitored during model optimization (Figure S2). The checkpoint corresponding to the best validation performance (Epoch 18) was selected as the final model for all subsequent analyses. The final model generated a continuous deep learning (DL)-derived risk score representing the probability of disease progression and was evaluated in two independent external validation cohorts. Quantification and statistical analysis Statistical analyses were performed using SPSS (version 20.0) and Python (version 3.9.1). Continuous variables were presented as mean ± standard deviation or median (interquartile range), and compared using the Student’s t test or Mann–Whitney U test as appropriate based on the Shapiro–Wilk test. Categorical variables were compared using the χ2 test or Fisher’s exact test. The deep learning (DL) model was developed using the training cohort and independently evaluated in two external validation cohorts. No data from the validation cohorts were used during model training or parameter tuning to avoid information leakage. Model discriminative performance was assessed using the area under the receiver operating characteristic curve (AUC), with comparisons performed using the DeLong test. Additional performance metrics included accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Cox proportional hazards regression analysis was performed to evaluate the association between the DL score and survival outcomes. Variables with a p value <0.05 in the univariate Cox regression analysis were entered into the multivariable Cox proportional hazards regression model. The variables included in the multivariable models differed among cohorts according to the results of the univariate analyses. Hazard ratios (HRs) and 95% confidence intervals (CIs) were calculated for all variables. Survival outcomes were analyzed using the Kaplan–Meier method and compared using the log rank test. The optimal cutoff value (0.501) was determined in the training cohort using the Youden index based on the Receiver operating characteristic (ROC) curve for 1-year progression status. Patients were then stratified into high- and low-risk groups using this cutoff, which was subsequently applied to the validation cohorts without modification. Restricted cubic spline (RCS) analysis was performed to evaluate the potential nonlinear association between the DL score and survival outcomes. The DL score was modeled as a continuous variable, and the reference value was set at the median DL score. Decision curve analysis (DCA) was performed to evaluate the clinical utility of the models. All statistical tests were two-sided, and p < 0.1 was considered statistically significant. Grad-CAM was applied to visualize regions contributing to model predictions for interpretability analysis. Data augmentation was applied only to the training cohort, whereas no augmentation was performed for either external validation cohort.