In a development that could reshape how doctors prepare lung cancer patients for treatment, researchers in China have built an artificial intelligence system that scans routine CT images of the entire lung — not just the visible tumor — to predict which patients with non-small cell lung cancer will achieve a complete pathological response to neoadjuvant immunotherapy. The study, published in BMC Cancer, analyzed data from 366 patients across two medical centers and demonstrated that a model combining tumor-focused and whole-lung radiomic features outperformed conventional approaches, offering clinicians a non-invasive tool for personalized treatment planning well before surgery begins.
Neoadjuvant immunotherapy, in which immune checkpoint inhibitors are administered before surgical resection, has transformed the management of locally advanced non-small cell lung cancer (NSCLC). When the treatment works exceptionally well, the tumor may vanish entirely under the microscope — a phenomenon called pathological complete response, or pCR — which is strongly associated with improved long-term survival and recurrence-free outcomes. The problem, however, is that pCR can only be confirmed definitively after the tumor has been surgically removed and examined by a pathologist. Patients who are unlikely to respond may undergo months of immunotherapy and its associated costs and toxicities for little benefit, while those most likely to benefit cannot be reliably identified in advance with current imaging or clinical assessments.
The research team, led by investigators at Tianjin Medical University Cancer Institute and Hospital in collaboration with Shandong Cancer Hospital and Institute, addressed this challenge using radiomics — a computational approach that extracts hundreds of quantitative features from medical images, capturing subtle patterns of texture, intensity, shape, and spatial relationships that are invisible to the human eye. Rather than focusing solely on the primary tumor, as most prior radiomic studies have done, the researchers deliberately analyzed the whole lung, reasoning that the tumor’s immunological microenvironment and the surrounding lung tissue might harbor additional predictive signals that a tumor-only analysis would miss.
The study enrolled 284 patients with stage II to IIIB NSCLC from the first center, treated between August 2018 and June 2025, who were randomly divided into a training set and an internal test set in a 7:3 ratio stratified by pCR status. An additional 82 patients from the second center, treated between April 2023 and March 2026, served as an external validation cohort — a critical step for confirming that the model’s performance generalizes beyond a single institution. All patients underwent baseline computed tomography before starting neoadjuvant immunotherapy, and radiomic features were extracted from two regions of interest: the primary tumor itself and the entire lung parenchyma.
To build the predictive models, the team employed LightGBM, a gradient-boosting decision tree algorithm known for its efficiency and accuracy with high-dimensional data. Two independent models were trained: one using only primary tumor (PT) features and one using only whole-lung (WL) features. The predicted probabilities from both models were then fused into a combined PTWL model. The results were striking. On the internal test set, the PT model achieved an area under the receiver operating characteristic curve (AUC) of 0.701, while the WL model reached 0.676 — comparable but individually modest. When the two were combined, however, the PTWL model’s AUC jumped to 0.788 internally and 0.787 on the external validation set, demonstrating that whole-lung information adds genuine predictive value beyond what the tumor alone can reveal.
Among the 366 patients in the overall cohort, whose median age was 63 years and who were predominantly male (317 of 366), 151 — or 41.3 percent — achieved a pathological complete response after neoadjuvant immunotherapy. The researchers further enhanced their model by incorporating carcinoembryonic antigen (CEA), a well-established serum tumor marker, into the PTWL signature. This combined model showed improved positive predictive value, reaching 0.769 in the internal test set and 0.636 in the external cohort, meaning that when the model predicted pCR, it was increasingly likely to be correct.
Perhaps the most clinically innovative aspect of the study is its hierarchical decision-making framework. Rather than relying on a single model output, the researchers designed a two-tiered system: the PTWL radiomic model performs initial risk stratification, sorting patients into preliminary risk categories, and then the combined model — incorporating both radiomics and CEA — serves as a secondary confirmation step for patients in specific risk tiers. This layered approach mirrors how clinicians actually make decisions, using multiple lines of evidence rather than a single test result, and it allows for more nuanced, tiered clinical decision-making than a one-size-fits-all prediction.
The framework’s clinical relevance was validated through survival analysis. Patients classified as low-risk by the hierarchical framework exhibited significantly prolonged recurrence-free survival, as demonstrated by Kaplan-Meier analysis with a log-rank P value below 0.05. This finding suggests that the model is not merely predicting a pathological endpoint but is also capturing information relevant to real-world oncological outcomes — a critical distinction for any tool intended to guide treatment decisions. Recurrence-free survival, the period after treatment during which no cancer returns, is one of the most meaningful endpoints for patients and oncologists alike.
From a technical standpoint, the study’s methodology reflects several best practices in radiomics research. The team assessed feature reproducibility using interclass correlation coefficients (ICC), a standard measure of how consistently features can be extracted across repeated measurements. Model discrimination was evaluated using receiver operating characteristic curves, and decision curve analysis (DCA) was employed to assess clinical utility — determining at which probability thresholds the model would provide net benefit if used in practice. The researchers also reported net reclassification improvements (NRI) and integrated discrimination improvements (IDI), sophisticated statistical measures that quantify how much better one model is at correctly classifying patients compared to another, providing rigorous evidence that the whole-lung integration genuinely improves prediction rather than merely adding complexity.
The implications of this work extend beyond NSCLC. The finding that whole-lung radiomic features — capturing the broader pulmonary environment, including emphysematous changes, inflammatory patterns, and parenchymal heterogeneity — contribute predictive information suggests that the tumor’s surroundings matter in immunotherapy response. This aligns with growing biological evidence that the immune landscape of the entire organ, not just the tumor itself, influences how checkpoint inhibitors perform. Patients with chronic obstructive pulmonary disease (COPD), for instance, may have a different immune contexture in their lungs that affects treatment response, and whole-lung analysis may implicitly capture such factors. As the authors conclude, their hierarchical framework facilitates individualized neoadjuvant treatment by reinforcing pCR identification, potentially sparing patients unlikely to benefit from unnecessary immunotherapy while confidently guiding those most likely to achieve complete pathological response toward the most effective treatment pathway. With further prospective validation, this approach could become a routine part of the pre-treatment workup for lung cancer patients worldwide.
Subject of Research: Whole-lung radiomics for predicting pathological complete response to neoadjuvant immunotherapy in non-small cell lung cancer
Article Title: Enhancing prediction of pathological complete response to neoadjuvant immunotherapy in non-small cell lung cancer by whole-lung radiomics
Article References: Cai, X., Yang, J., E, J., Wang, D., Qu, F., Wang, Q., Ning, Z., Liu, F., Zhang, Y., Gao, B., Sun, B., & Liu, Y. (2026). Enhancing prediction of pathological complete response to neoadjuvant immunotherapy in non-small cell lung cancer by whole-lung radiomics. BMC Cancer. https://doi.org/10.1186/s12885-026-16939-7
Image Credits: AI Generated
DOI: 10.1186/s12885-026-16939-7
Keywords: non-small cell lung cancer, neoadjuvant immunotherapy, radiomics, pathological complete response, artificial intelligence, LightGBM, CT imaging, CEA, recurrence-free survival, predictive medicine, tumor microenvironment, personalized oncology
News Source: Nathaniel Bowman. (October 18, 2026). AI Reads the Whole Lung to Predict Which Cancer Patients Will Benefit From Immunotherapy. Scienmag.
