---
title: "CT-based deep learning predicts ICI-plus-chemotherapy response after first-generation EGFR-TKI res"
id: "frontiers-in-immunology-15-predicting-response-to-immune-checkpoint-inhibitor-plus-chemotherapy-in-egfr"
canonical_url: "https://medichelpline.com/clinical-feed/frontiers-in-immunology-15-predicting-response-to-immune-checkpoint-inhibitor-plus-chemotherapy-in-egfr"
content_type: "clinical_feed_article"
specialty: "Oncology"
source_name: "Frontiers in Immunology"
source_url: "https://www.frontiersin.org/articles/10.3389/fimmu.2026.1760264"
published_at: "2026-07-17T00:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# CT-based deep learning predicts ICI-plus-chemotherapy response after first-generation EGFR-TKI res
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/frontiers-in-immunology-15-predicting-response-to-immune-checkpoint-inhibitor-plus-chemotherapy-in-egfr
- **Specialty:** [Oncology](https://medichelpline.com/clinical-feed/oncology.md)
- **Primary Source:** Frontiers in Immunology
- **Source URL:** [Original Journal Publication](https://www.frontiersin.org/articles/10.3389/fimmu.2026.1760264)
- **Published At:** 2026-07-17T00:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- This multicenter retrospective study developed and externally validated CT-based deep learning models to predict objective response to **ICI-chemotherapy** in patients with **EGFR-mutant lung adenocarcinoma** who progressed on first-generation EGFR-TKIs and lacked the T790M resistance mutation. - The final analyzable cohort comprised 490 patients from three centers: a training cohort (n = 326) and two independent validation cohorts (n = 70 and n = 94). All patients had post-resistance, pre-immunotherapy contrast-enhanced CT imaging and molecular profiling confirming absence of T790M. - Models compared included 2D, 2.5D, and 3D ResNet-101 architectures using axial, sagittal, and coronal tumor-centered inputs; the 2.5D approach stacked five adjacent slices per view to supply limited through-plane context while leveraging 2D pretrained weights. - Lesions were segmented by two blinded observers with excellent interobserver reproducibility (ICC median 0.998). Training used ImageNet-initialized backbones for 2D/2.5D; the 3D model was trained from scratch. No clinical-only or combined clinical-imaging models were constructed due to incomplete harmonization of some clinical variables. - The primary endpoint for model training was objective response (CR/PR vs SD/PD by RECIST 1.1); PFS was used for post-hoc survival stratification. Classification thresholds were derived in the training set using Youden index and locked for external validation. - Performance: the **2.5D axial model** achieved AUCs of **0.885** (training), **0.819** (validation 1), and **0.863** (validation 2). Model outputs were also evaluated for PFS-based risk stratification; overall ORR across cohorts was 34.1%. - The study highlights a noninvasive imaging strategy to predict ICI benefit in a challenging post-TKI resistance population but recognizes the need for prospective validation, formal benchmarking versus clinical variables, and harmonized prospective data before clinical implementation. - Limitations explicitly reported include retrospective design, incomplete clinical variable harmonization preventing clinical-only or combined models, and absence of prospective validation or external clinical benchmarking in this manuscript.
## Clinical Analysis & Structured Key Points
About us All journals All articles Submit your research Search Login Frontiers in Immunology Sections Articles Research Topics Editorial board About journal Published in Frontiers in Immunology Cancer Immunity and Immunotherapy 7 impact factor 11.3 citescore Part of a Research Topic Evolving non-invasive biomarkers: Chemokines and multi-omic profiling in NSCLC immunotherapy response 3691 views 6 articles Reviewers Reviewed by Zhirui Zeng SYED NURUL HASAN Outline Abstract 1 Introduction 2 Materials and methods 3 Results 4 Discussion Data availability statement Ethics statement Author contributions Funding Conflict of interest Generative AI statement Publisher’s note Supplementary material Abbreviations References Figures and Tables Figure 1 View in article Figure 2 View in article Figure 3 View in article Figure 4 View in article Figure 5 View in article Table 1 Baseline characteristics across the training cohort and two validation cohorts. View in article ORIGINAL RESEARCH article Front. Immunol., 17 July 2026 Sec. Cancer Immunity and Immunotherapy Volume 17 - 2026 | https://doi.org/10.3389/fimmu.2026.1760264 Predicting response to immune checkpoint inhibitor plus chemotherapy in EGFR-mutant lung adenocarcinoma following first-generation TKI resistance: a multicenter deep learning study Shuai Qie Y S Yasong Shi J L Jingyun Li S J Sicong Jia X Y Xiaoping Yin * Department of Radiation Oncology, Hebei University Affiliated Hospital, Baoding, Hebei, China Article metrics View details 67 Views Abstract Background: Patients with epidermal growth factor receptor (EGFR)-mutant lung adenocarcinoma who develop resistance to first-generation tyrosine kinase inhibitors (TKIs) without a T790M mutation face a therapeutic dilemma with limited and suboptimal options. Methods: In this multicenter retrospective study, the final analyzable modeling cohort included 490 patients with complete eligible CT imaging and outcome labels, comprising a training cohort of 326 patients, validation cohort 1 of 70 patients, and validation cohort 2 of 94 patients. Model performance was evaluated using AUC, decision curve analysis, and PFS stratification analyses. Results: The 2.5D axial model achieved AUCs of 0.885, 0.819, and 0.863 in the training cohort, validation cohort 1, and validation cohort 2, respectively, based on the locked source prediction files. Conclusion: A CT-based 2.5D deep learning model showed promising performance for treatment-response prediction after EGFR-TKI resistance, but prospective validation and clinical-variable benchmarking remain necessary before clinical implementation. 1 Introduction Lung cancer remains the leading cause of cancer-related mortality globally, with EGFR mutations being the most prevalent oncogenic drivers in adenocarcinoma (1). While EGFR-TKIs like Osimertinib have revolutionized first-line treatment, acquired resistance invariably develops, creating a critical therapeutic challenge with limited and suboptimal subsequent options (2, 3). ICIs have reshaped the landscape of advanced non-small cell lung cancer (NSCLC) without actionable drivers. However, their efficacy in EGFR-mutant NSCLC, particularly after TKI failure, remains modest and unpredictable, compounded by the risk of hyperprogression (4, 5). Current biomarkers for guiding ICI therapy in this setting are inadequate. PD-L1 tumor proportion score demonstrates limited predictive value and significant heterogeneity (6), while obtaining post-resistance tumor tissue for advanced profiling (e.g., tumor mutational burden, tumor microenvironment analysis) is often impractical due to its invasive nature, sampling bias, and patient ineligibility (7). Consequently, there is an urgent, unmet need for a non-invasive tool to predict ICI benefit in this patient population. Deep learning has emerged as a powerful technique for decoding complex information from medical images (8, 9). We hypothesize that the dynamic biological evolution of tumors during the development of TKI resistance—including alterations in the tumor immune microenvironment—imparts discernible, albeit subtle, signatures on routine CT scans. These latent imaging phenotypes may encapsulate critical information predictive of subsequent response to immunotherapy (10, 11). Existing studies are primarily limited to predicting static genetic alterations using radiomics or deep learning, as consolidated by a systematic review on EGFR mutation prediction in NSCLC (12, 13) or focus on ICI response in TKI-naïve populations (14–16). A significant gap exists in leveraging advanced deep learning to directly predict ICI efficacy specifically in the post-TKI resistance setting. Furthermore, many approaches rely on multi-step, handcrafted feature engineering, which may introduce bias and fail to capture the most salient patterns. To address this, we propose that an end-to-end deep learning model can directly decode these complex visual patterns from baseline CT images acquired after EGFR-TKI resistance to predict immunotherapy outcomes accurately. This study aims to develop and validate such a model using a large, multi-center retrospective cohort. Our framework takes post-resistance, pre-immunotherapy CT images as direct input to predict objective response to ICI-chemotherapy and to evaluate model-derived PFS risk stratification. 2 Materials and methods 2.1 Study participants The patient selection process is shown in Figure 1. The initially identified screening cohort consisted of 750 patients from three centers. After application of predefined inclusion and exclusion criteria, 260 patients were excluded, resulting in a final analyzable cohort of 490 patients with complete eligible CT imaging and treatment response labels. The reasons for exclusion were as follows: (1) missing follow-up or unavailable treatment response labels (n = 32), (2) poor-quality or incomplete CT imaging unsuitable for quantitative analysis (n = 74), and (3) failure to meet predefined eligibility criteria, including absence of sensitizing EGFR mutation, presence of T790M or other driver mutations, prior systemic therapy before EGFR-TKI initiation, or incomplete clinical records (n = 154). To ensure transparency and consistency, cohort nomenclature was standardized throughout the revised manuscript, figures, and Supplementary Materials. The final analyzable cohort was divided into a training cohort (n = 326), validation cohort 1 (n = 70), and validation cohort 2 (n = 94). The training cohort was used for model development, including model fitting, model selection, and threshold derivation. Validation cohort 1 and validation cohort 2 were derived from institutionally independent centers and therefore served as two independent external validation cohorts. Both validation cohorts were used exclusively for independent performance evaluation. No data from either validation cohort were used for model training, hyperparameter tuning, threshold selection, threshold derivation, or model selection. Figure 1 Study participant flowchart. 2.2 Molecular profiling and treatment pathway justification To ensure the enrolled population was representative of patients for whom ICI-chemotherapy is a standard option, all patients were required to have undergone molecular profiling following resistance to first-generation EGFR-TKIs. Patients were included only if there was no evidence of the EGFR T790M resistance mutation in their medical records, making them ineligible for third-generation EGFR-TKIs (e.g., Osimertinib). The exclusion of other driver mutations was also verified as per the exclusion criteria. Molecular profiling was primarily performed using tissue-based next-generation sequencing (NGS). When tissue was unavailable or insufficient, liquid biopsy (circulating tumor DNA analysis) was used as an alternative. This stringent patient selection guarantees that the study cohort accurately reflects the clinical scenario of patients with EGFR-mutant lung adenocarcinoma who have exhausted targeted therapy options and are candidates for ICI-chemotherapy. 2.3 Clinical and pathological data collection Comprehensive clinical and pathological data were retrospectively collected for all enrolled patients. The variables included age, sex, smoking history, Eastern Cooperative Oncology Group (ECOG) performance status, pretreatment clinical stage, tumor location, specific EGFR mutation subtype (exon 19 deletion or L858R), treatment regimen, and subsequent systemic therapy after ICI-chemotherapy where available. PD-L1 tumor proportion score was also collected when it had been tested in routine clinical care; because PD-L1 testing was not uniformly available across all participating centers, PD-L1-based analyses were considered exploratory subset analyses. 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). 2.4 Pretreatment evaluation Prior to the initiation of immune-checkpoint inhibitor combined with chemotherapy (ICI-chemotherapy), a comprehensive pretreatment evaluation was performed for all patients to ensure they were suitable candidates and to establish a baseline for response assessment. 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 obligatory. Laboratory tests included a complete blood count, comprehensive metabolic panel, and screening for hepatitis B and C serology. This standardized workup ensured adequate organ function and no uncontrolled comorbidities. 2.5 Treatment regimens Following confirmed disease progression after first-generation EGFR-TKI therapy, enrolled patients received ICI plus chemotherapy as subsequent systemic therapy. Chemotherapy regimens included platinum-based combinations with pemetrexed, paclitaxel, or other physician-selected regimens, and immune checkpoint inhibitors mainly included PD-1 or PD-L1 inhibitors. Treatment was administered according to institutional practice and contemporary clinical guidelines. 2.6 Response assessment and follow-up Tumor response was evaluated by board-certified radiologists according to the Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1. The radiologists were blinded to model predictions and cohort allocation during response assessment. Radiographic assessments were conducted every two cycles, approximately every 6 weeks, during the first 6 months of ICI-chemotherapy and every 3 months thereafter until disease progression, death, or the last available follow-up. The primary classification endpoint for model development was objective response after ICI-chemotherapy. Response was defined as the best overall response achieved during ICI-chemotherapy and was classified as complete response (CR), partial response (PR), stable disease (SD), or progressive disease (PD) according to RECIST 1.1. Objective response rate (ORR) was defined as the proportion of patients who achieved CR or PR. In the locked label file, label = 1 denoted response-positive disease, corresponding to CR or PR, whereas label = 0 denoted response-negative disease, corresponding to SD or PD. Progression-free survival (PFS) was used as the time-to-event endpoint for survival stratification. PFS was defined as the time interval from initiation of ICI-chemotherapy to the first documented radiological progression or death from any cause. Patients without progression or death were censored at the date of their last radiographic assessment. Overall survival was not included in the survival analyses reported in this manuscript because it may be substantially confounded by subsequent lines of therapy after disease progression. 2.7 CT acquisition and preprocessing Figure 2 depicts the schema of the present study. Lesion segmentation was performed independently by two observers who were blinded to treatment outcomes. Interobserver reproducibility was assessed using the available 30-case subset with paired feature extraction files. ICC values ranged from 0.973 to 1.000, with a median ICC of 0.998, supporting excellent reproducibility. Cases with segmentation disagreement were resolved by consensus review. Figure 2 Schematic overview of the deep learning study workflow. 2.8 2D Resnet-101 model development A two-dimensional (2D) ResNet-101 model was developed to predict objective response to ICI-chemotherapy using CT images from three orthogonal views: axial, sagittal, and coronal. For each view, the tumor-centered slice with the largest tumor area was selected according to the segmented region of interest (ROI). The ROI patch was cropped and resized to 256 × 256 pixels before being used as model input. Three independent view-specific ResNet-101 models were constructed for the axial, sagittal, and coronal views. The ResNet-101 backbone was initialized using ImageNet-pretrained weights and fine-tuned on the training cohort. Data augmentation was applied during training, including random horizontal and vertical flipping and center cropping. The final fully connected layer was replaced with a task-specific binary classification layer, and the model output represented the predicted probability of response-positive disease. Binary cross-entropy loss was used for the ORR classification task. Model parameters were optimized using the Adam optimizer with a batch size of 64, and dropout regularization was applied to reduce overfitting. After training the three view-specific models, feature representations from the penultimate layers were fused and passed through a final classifier to generate the overall 2D prediction score for objective response. 2.9 2.5D Resnet-101 model development In this study, the term 2.5D refers to a multi-channel pseudo-volumetric approach in which five adjacent CT slices are stacked as input channels and processed by a 2D convolutional neural network backbone. This strategy was designed to incorporate limited through-plane spatial context while retaining the computational efficiency and transfer-learning advantages of 2D CNNs. For each orthogonal view, including axial, sagittal, and coronal views, the central slice with the largest tumor area was first identified. Two adjacent slices on each side of the central slice were then selected at 1-mm intervals, resulting in five consecutive slices per view. These five slices were stacked as five input channels to construct the 2.5D input. For each view, the minimum bounding rectangle encompassing the tumor region across all five slices was defined as the ROI, cropped, and resized to 256 × 256 pixels. Three independent 2.5D ResNet-101 models were trained for the axial, sagittal, and coronal views. The ResNet-101 backbone was initialized using ImageNet-pretrained weights and fine-tuned on the training cohort. The first convolutional layer was modified to accommodate five-channel input. The remaining network architecture followed the standard ResNet-101 design with residual skip connections. Data augmentation strategies were consistent with those used for the 2D models, including random flipping and center cropping. The final fully connected layer was replaced with a binary classification layer, and the model was trained to output the predicted probability of response-positive disease after ICI-chemotherapy. Binary cross-entropy loss was used as the classification loss. Model optimization was performed using the Adam optimizer with an initial learning rate of 0.001 and a batch size of 64. Dropout regularization was incorporated to reduce overfitting. The 2.5D model therefore learned an enriched tumor representation from adjacent slices and multiple anatomical planes for end-to-end prediction of objective response. 2.10 3D Resnet-101 model development To compare the 2D and 2.5D approaches with a volumetric deep learning strategy, a three-dimensional (3D) ResNet-101 model was also developed. In 3D data processing, tumor ROIs were encapsulated within bounding cubes and uniformly resampled to 96 × 96 × 96 voxels using linear interpolation. Data augmentation was applied by mirror flipping along the X, Y, and Z axes. Because directly transferable ImageNet-pretrained weights are not readily available for standard 3D medical imaging models, the 3D ResNet-101 model was initialized from scratch and trained on the training cohort. The final fully connected layer was replaced with a binary classification layer to predict response-positive disease after ICI-chemotherapy. Binary cross-entropy loss was used for model training. The model was trained for 200 epochs using the Adam optimizer. The 3D model output represented the predicted probability of objective response. No survival-specific loss function or Cox proportional hazards layer was used in the deep learning models. PFS was evaluated separately as a time-to-event endpoint for risk stratification based on the model output score, as described in the statistical analysis section. 2.11 Model training, class imbalance, and threshold selection All model training, hyperparameter selection, and threshold derivation were performed using the training cohort only. Validation cohort 1 and validation cohort 2 were not used for model training, hyperparameter tuning, threshold selection, threshold derivation, or model selection. In the training cohort, the response-positive and response-negative groups included 118 and 208 patients, respectively, indicating moderate class imbalance. No oversampling, undersampling, or class weighting strategy was applied. The class distribution was reported to support transparent interpretation of threshold-dependent performance metrics. For each model, the classification threshold was determined in the training cohort using the Youden index and then applied unchanged to validation cohort 1 and validation cohort 2. Thresholds were not re-optimized in either validation cohort. 2.12 Statistical analysis Statistical analyses were performed using SPSS 20.0 and Python 3.9.1. Continuous variables were compared using t-tests or Mann-Whitney U tests, and categorical variables were analyzed using chi-square or Fisher’s exact tests. Model performance was evaluated using AUC, accuracy, sensitivity, specificity, PPV, NPV, and F1 score. For each architecture, the classification threshold was selected in the training cohort using the Youden index and then applied unchanged to validation cohort 1 and validation cohort 2. PFS was analyzed as the time-to-event endpoint for survival stratification. Due to incomplete availability and limited harmonization of some clinical variables across centers, formal clinical-only and combined clinical-imaging predictive models were not constructed in the present study. This issue is acknowledged as a limitation and should be addressed in future prospective studies. 3 Results 3.1 Clinical characteristics The final analyzable cohort included 490 patients, comprising 326 patients in the training cohort, 70 patients in validation cohort 1, and 94 patients in validation cohort 2. According to RECIST-defined objective response, the response-positive/response-negative distribution was 118/208 in the training cohort, 20/50 in validation cohort 1, and 29/65 in validation cohort 2, corresponding to ORRs of 36.2%, 28.6%, and 30.9%, respectively. The overall ORR was 34.1% among the 490 patients. Baseline clinical characteristics across the three cohorts are summarized in Table 1. Most baseline variables, including age, sex, T stage, N stage, primary tumor site, smoking status, comorbidities, EGFR subtype, metastatic status, chemotherapy regimen, and immune checkpoint inhibitor category, were generally comparable across cohorts. However, dr
## Related Clinical Research

- [Kennedy renews attacks on vaccines at Children’s Health Defense conference](https://medichelpline.com/clinical-feed/stat-news-1-kennedy-renews-attacks-on-vaccines-at-children-s-health-defense-conference.md)
- [Intratumoral Mycobacterium abscessus and cytidine deaminase mutagenesis in non-small cell lung can](https://medichelpline.com/clinical-feed/pubmed-42675040.md) (DOI: 10.1038/s41392-026-02878-z)
- [Acquired immune-mediated TTP after ivonescimab chemoimmunotherapy with cholestatic liver injury an](https://medichelpline.com/clinical-feed/frontiers-in-immunology-18-acquired-immune-mediated-thrombotic-thrombocytopenic-purpura-with-severe.md)
- [Reproducibility of DECT-derived Perfusion Blood Volume and 4DCT Ventilation Imaging in NSCLC](https://medichelpline.com/clinical-feed/pubmed-42716107.md) (DOI: 10.1088/1361-6560/aea51b)
- [LongPhase-TO: tumor-only long-read method for somatic haplotype reconstruction and variant recalib](https://medichelpline.com/clinical-feed/biorxiv-8-somatic-haplotype-reconstruction-and-variant-recalibration-from-tumor-only-long.md)

## Navigation
- [← Back to Oncology Feed](https://medichelpline.com/clinical-feed/oncology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.