AI-Powered CT Imaging Emerges as a Game-Changer in Predicting Immunotherapy Response in Advanced NSCLC
In the rapidly evolving landscape of oncology, immune checkpoint inhibitors (ICIs) have redefined the treatment paradigm for advanced non-small-cell lung cancer (NSCLC). Offering durable responses and improved survival for a subset of patients, ICIs have become a cornerstone of first-line therapy—particularly for those lacking actionable driver mutations. Yet, a persistent clinical dilemma remains: only 15% to 20% of patients derive long-term benefit from these therapies, while others experience no response or, worse, hyperprogressive disease. This stark variability underscores an urgent need for reliable, non-invasive biomarkers capable of predicting treatment outcomes before or early during therapy.
Enter artificial intelligence (AI). Specifically, AI-driven analysis of routine computed tomography (CT) scans is emerging as a powerful, cost-effective, and scalable solution to this challenge. By leveraging radiomics and deep learning, researchers can extract hundreds—even thousands—of quantitative features from standard-of-care imaging that are invisible to the human eye. These features, when processed through sophisticated machine learning models, reveal hidden patterns correlated with tumor biology, immune microenvironment, and clinical outcomes. Recent studies demonstrate that such AI models not only predict who will respond to immunotherapy but also identify patients at risk of hyperprogression—a rare but devastating complication.
This shift represents more than a technical advance; it signals a fundamental reimagining of medical imaging. No longer just a tool for anatomical assessment, CT is becoming a dynamic data source capable of informing precision oncology decisions in real time.
The Limitations of Current Imaging Criteria
Traditionally, oncologists have relied on the Response Evaluation Criteria in Solid Tumors (RECIST) to assess treatment response. While RECIST 1.1 provides a standardized framework for measuring tumor size changes, it falls short in the context of immunotherapy. Unlike chemotherapy or targeted therapy, which typically cause rapid tumor shrinkage, immunotherapies can trigger complex immune-mediated responses that defy conventional interpretation.
For instance, some patients exhibit “pseudoprogression”—an initial increase in tumor size or appearance of new lesions due to immune cell infiltration, followed by subsequent regression. Others show “mixed responses,” where some lesions shrink while others grow. In extreme cases, a small but significant subset of patients experience hyperprogression, defined by a dramatic acceleration in tumor growth kinetics shortly after ICI initiation. These atypical patterns can lead to premature discontinuation of potentially effective therapy or, conversely, continued treatment in patients who are actively deteriorating.
To address these challenges, the iRECIST guidelines were introduced in 2017. They add categories like immune unconfirmed progressive disease (iUPD) and immune confirmed progressive disease (iCPD), requiring confirmatory imaging 4–8 weeks after an initial progression signal. While iRECIST improves data consistency in clinical trials, it remains reactive rather than predictive. It does not help clinicians decide before treatment whether a patient is likely to benefit—or suffer harm—from ICIs.
Moreover, even with iRECIST, interpretation remains subjective and limited to macroscopic changes. It offers no insight into the underlying tumor-immune interactions that determine therapeutic success or failure.
Radiomics: Turning Pixels into Prognostic Power
Radiomics bridges this gap by converting medical images into mineable high-dimensional data. The process begins with segmentation of tumor regions on baseline CT scans—either manually by radiologists or increasingly via automated AI tools. From these regions, algorithms extract a vast array of features describing tumor intensity, texture, shape, and heterogeneity. These features are then fed into machine learning models trained to correlate imaging phenotypes with clinical outcomes.
One landmark study led by Sun et al. demonstrated the feasibility of this approach across multiple cancer types, including NSCLC. Using pre-treatment contrast-enhanced CT scans from 135 patients in a phase I trial of PD-1/PD-L1 inhibitors, the team developed a radiomic signature predictive of CD8+ T-cell infiltration—a key determinant of immunotherapy response. The model successfully distinguished “immune-inflamed” tumors (rich in T cells) from “immune-desert” tumors (lacking T cells). More importantly, patients with high radiomic scores had significantly longer median overall survival (24.3 months vs. 11.5 months) and were more likely to achieve durable clinical benefit.
Critically, this signature was validated across three independent cohorts, including one with known tumor mutational burden (TMB)—another established biomarker for ICI response. The consistency of performance across diverse datasets suggests that radiomics captures biologically meaningful signals that transcend individual patient or institutional variables.
Deep Learning Goes Beyond Handcrafted Features
While traditional radiomics relies on pre-defined mathematical features, deep learning takes a more holistic approach. Convolutional neural networks (CNNs), particularly 3D architectures like DenseNet, can automatically learn hierarchical representations directly from raw CT volumes without manual feature engineering.
He et al. applied this technique to 327 NSCLC patients with known TMB status. Their deep learning model, termed the TMB Radiomic Biomarker (TMBRB), outperformed conventional radiomic models in distinguishing high-TMB from low-TMB tumors. When used to stratify patients by predicted ICI response, TMBRB identified groups with significantly different progression-free survival (PFS) and overall survival (OS). The hazard ratio for OS was 0.54, indicating a 46% reduction in risk of death for the low-risk group.
Even more compelling, the model’s predictive power increased when combined with clinical variables like Eastern Cooperative Oncology Group (ECOG) performance status. This synergy between imaging and clinical data exemplifies the potential of multimodal AI systems to deliver truly personalized risk assessment.
Capturing Dynamic Changes with Delta Radiomics
Baseline imaging, while informative, provides only a static snapshot. Tumors evolve under therapeutic pressure, and early changes may be more predictive than pre-treatment characteristics alone. This insight has given rise to “delta radiomics”—the analysis of changes in radiomic features between serial scans.
Khorrami et al. pioneered this approach in NSCLC by comparing pre- and early-on-treatment CT scans. They extracted 618 features from both the tumor core and its surrounding microenvironment, then computed the relative change (delta) for each feature. After rigorous filtering for reproducibility—using test-retest CT scans from the publicly available RIDER dataset—they identified eight robust delta features predictive of immunotherapy response.
The resulting model achieved area under the curve (AUC) values of 0.88, 0.85, and 0.81 in training and two validation cohorts—consistently outperforming models based on baseline features alone. Notably, the most predictive features originated not from the tumor itself but from the peritumoral region, suggesting that the immune response extends beyond visible boundaries.
Even more intriguingly, these delta features correlated with the density of tumor-infiltrating lymphocytes (TILs) on histopathology slides. Using computational pathology algorithms to segment lymphocyte nuclei based on size, shape, and staining intensity, the researchers found a statistically significant association between peritumoral Gabor filter delta features and TIL density. This cross-modality validation strengthens the biological plausibility of delta radiomics as a surrogate for immune activity.
Identifying Hyperprogression Before It’s Too Late
Perhaps the most urgent application of AI in this space is the early detection of hyperprogression. Defined by a ≥2-fold increase in tumor growth rate after ICI initiation, hyperprogression affects an estimated 10–20% of NSCLC patients and is associated with median survival of less than two months. Currently, no validated biomarker exists to identify these patients prospectively.
Vaidya et al. tackled this problem by analyzing baseline CT scans from 109 advanced NSCLC patients treated with PD-1/PD-L1 inhibitors. They extracted 198 radiomic features, including quantitative vessel tortuosity (QVT)—a measure of abnormal tumor vasculature. Using random forest feature selection, they identified three key predictors: one peritumoral texture feature and two QVT metrics.
The resulting model achieved remarkable performance: AUCs of 0.85 in the training set and 0.96 in the test set. Kaplan-Meier analysis confirmed that patients classified as high-risk for hyperprogression had significantly worse overall survival (hazard ratio = 2.66). If validated prospectively, such a tool could spare high-risk patients from ineffective—and potentially harmful—therapy, redirecting them toward alternative strategies like chemotherapy or clinical trials.
Multimodal Integration: The Next Frontier
While CT-based AI shows immense promise, combining it with other imaging modalities may unlock even greater accuracy. Mu et al. explored this by analyzing baseline PET/CT scans from 194 stage IIIB–IV NSCLC patients. They extracted radiomic features separately from the CT component, the FDG-PET component, and the fused PET/CT image, then integrated them into a multiparametric radiomic signature (mpRS).
The mpRS outperformed any single-modality model in predicting durable clinical benefit (DCB), with AUCs of 0.89, 0.86, and 0.86 across training, retrospective, and prospective test sets. When combined with clinical factors like histology and ECOG score, the final model achieved C-indices of up to 0.83 for OS prediction—rivaling or exceeding many genomic biomarkers in practical utility.
Decision curve analysis further confirmed that the integrated model provided the highest net clinical benefit across a wide range of risk thresholds, making it suitable for real-world decision support.
Toward Clinical Implementation
Despite these advances, several challenges remain before AI-powered CT analysis becomes routine in oncology clinics. First, model generalizability is critical. Most studies to date use retrospective, single-institution data. Prospective, multicenter trials are needed to validate performance across diverse scanners, protocols, and populations.
Second, standardization is essential. Variations in CT acquisition parameters (e.g., slice thickness, contrast timing) can significantly affect radiomic features. Efforts like the Image Biomarker Standardization Initiative (IBSI) aim to harmonize feature definitions, but widespread adoption is still pending.
Third, regulatory and workflow integration hurdles must be addressed. AI tools must be embedded seamlessly into radiology information systems (RIS) and electronic health records (EHR), with clear interpretability for clinicians. Black-box models, no matter how accurate, will struggle to gain trust without transparent decision logic.
Nevertheless, the trajectory is clear. With its non-invasive nature, universal availability, and rich data content, CT imaging—supercharged by AI—stands poised to become a central pillar of precision immuno-oncology. Unlike tissue biopsies, which are invasive, spatially limited, and difficult to repeat, CT scans are routinely performed throughout the treatment course, enabling dynamic monitoring and adaptive therapy.
The Future: From Prediction to Prescription
Looking ahead, the ultimate goal is not just to predict response but to guide therapeutic decisions. Imagine a scenario where, at the time of ICI initiation, an AI system analyzes the patient’s baseline CT and outputs a risk score for non-response, hyperprogression, and long-term survival. Based on this, the oncologist might choose to combine ICI with chemotherapy for high-risk patients, or enroll them in a trial of novel immunomodulators.
Moreover, as AI models incorporate longitudinal imaging, they could detect early signs of resistance—weeks before tumor growth becomes apparent on RECIST—allowing timely switch to alternative regimens. In this vision, imaging becomes not just a diagnostic tool but a continuous feedback loop for adaptive cancer therapy.
The convergence of radiology, oncology, and artificial intelligence is reshaping how we understand and treat cancer. For patients with advanced NSCLC, this integration offers hope for smarter, safer, and more effective use of immunotherapy—ensuring the right treatment reaches the right patient at the right time.
Author: Liu Ying, Li Qian, Zhang Yuwei, Ye Zhaoxiang
Affiliation: Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer; Key Laboratory of Cancer Prevention and Therapy, Tianjin; Tianjin’s Clinical Research Center for Cancer, Tianjin 300060, China
Journal: International Journal of Medical Radiology
DOI: 10.19300/j.2021.Z18620