Immunotherapy selection depends on PD-L1 status. In non-small cell lung cancer (NSCLC), which accounts for approximately 85% of all lung cancer diagnoses, the expression of programmed death ligand 1 (PD-L1) is the primary biomarker for selecting patients who will benefit from immune checkpoint inhibitor (ICI) therapy. Patients with high PD-L1 expression can receive these drugs as first-line monotherapy, avoiding more toxic platinum-based chemotherapy regimens.
The limitations of tissue biopsy. PD-L1 expression is currently assessed by immunohistochemistry (IHC) on biopsy samples. However, biopsy is invasive and can fail due to tumor inaccessibility, extensive emphysema, patient performance status, or risk of complications such as pneumothorax, bleeding, and infection. Additionally, PD-L1 expression can be spatially heterogeneous within the tumor and can change over time with treatment, making a single biopsy sample potentially unrepresentative.
Radiomics as a non-invasive alternative. Radiomics extracts hundreds of quantitative features from standard medical images -- capturing texture, intensity patterns, and shape characteristics invisible to the human eye. The central premise is that these imaging features encode underlying tumor biology, including molecular marker expression, enabling a form of non-invasive digital biopsy.
Why PET/CT is the right imaging modality. 18F-FDG-PET/CT is already recommended by NCCN guidelines for all NSCLC patients for primary staging. PD-L1 expression is biologically linked to glucose metabolism through the Akt/mTOR/HIF-1alpha signaling axis, which upregulates glycolytic enzymes and glucose transporter GLUT-1. This mechanistic link suggests that FDG uptake patterns captured by PET/CT may reflect PD-L1 expression levels in the tumor.
Retrospective cohort at Semmelweis University. The study analyzed 18F-FDG-PET/CT scans from 148 NSCLC patients collected between March 2017 and March 2021 at the Department of Nuclear Medicine, Medical Imaging Centre, Semmelweis University, Budapest. After applying exclusion criteria (unknown PD-L1 status, incomplete imaging, peritumoral atelectasis), 105 patients were included in the final radiomic analysis.
Patient characteristics and PD-L1 classification. The cohort included 72 adenocarcinoma (ACC) and 33 squamous cell carcinoma (SCC) patients. PD-L1 positivity was defined as 1% or greater tumor proportion score by IHC, classifying 64 patients as PD-L1 positive and 41 as negative. The cohort reflected a real-world population of mostly elderly former or active smokers across all disease stages.
Standardized PET/CT acquisition protocol. Scans were performed on a GE Discovery IQ 5 PET/CT system following EANM procedure guidelines. Patients fasted for at least 6 hours and received 2.5 MBq/kg of pharmaceutical-grade 18F-FDG intravenously, with an uptake time of 55 to 65 minutes. Images were reconstructed using Bayesian penalized likelihood (Q.Clear) for PET and included both a low-dose CT for attenuation correction and a full-inspiration diagnostic chest CT.
Image segmentation approach. All primary tumor regions of interest (ROIs) were manually delineated on PET images by an experienced nuclear medicine physician using 3D isocontour drawing in InterView FUSION software. Because PET and CT have different spatial resolutions and respiratory motion can cause misalignment, segmentation masks were resampled and where necessary manually translated in 3D Slicer to correctly overlay the tumor on CT images.
Multimodal radiomic feature extraction. Features were extracted from both PET and CT images using the PyRadiomics Python package. CT features were extracted from the original image and from wavelet-transformed versions (eight decomposition combinations: LLL, LLH, LHL, LHH, HLL, HLH, HHL, HHH) and Laplacian of Gaussian (LoG) transforms at sigma values of 1, 2, 3, and 4 mm. All images were normalized and resampled to 1x1x1 mm voxels with a bin width of 25.
Three-stage feature reduction pipeline. Given the large number of extracted features relative to the 105-patient sample, a three-stage reduction was applied: first, Mann-Whitney U tests excluded features that did not show significant distributional differences between PD-L1 positive and negative groups; second, LASSO regression (with optimal regularization parameter alpha determined by ten-fold cross-validation) further narrowed features by shrinking non-informative coefficients to zero; and third, Spearman rank-order correlation removed highly correlated feature pairs, retaining only the one with higher correlation to PD-L1 status.
Data leakage prevention. The dataset was split 7:3 into training and test sets using stratified splitting to maintain the same proportion of PD-L1 positive and negative patients in both groups. All feature selection steps were performed exclusively on the training set to prevent data leakage into the evaluation.
Final model construction. After feature selection, a linear logistic regression classifier was built using the retained features. Two cases were evaluated, differing in how the PET-to-CT mask translation was handled (direct transfer versus manually corrected translation), to assess the impact of segmentation quality on model performance.
SUV differences between PD-L1 groups. PD-L1 positive tumors showed significantly higher FDG uptake than negative tumors, with median SUVmax of 15.3 vs 12.3 (p=0.00039) and median SUVpeak of 12.8 vs 9.2 (p=0.00044). This confirms the biological link between PD-L1 expression and enhanced glucose metabolism and provides a basis for PET-based radiomic prediction.
Best model performance. The optimal model (Case 2, with manually corrected segmentation) achieved an AUC of 0.783 (95% CI 0.625-0.942) on the test set, with accuracy 81.25%, sensitivity 90.00%, specificity 66.67%, positive predictive value 81.81%, and negative predictive value 80.00%. The high sensitivity of 90% is clinically important because it means most true PD-L1 positive patients are correctly identified for immunotherapy eligibility.
Selected features and their meaning. Five features were selected in each case. For Case 2, they included wavelet-LHL GLSZM Zone Percentage and log-sigma ngtdm Busyness (CT texture features capturing spatial complexity), original firstorder Total Energy and Interquartile Range (PET intensity features), and GLDM Dependence Non-Uniformity Normalized (CT uniformity). The combination of PET intensity and CT texture features highlights the complementary value of multimodal imaging.
Segmentation quality impacts results. Case 1, using direct mask transfer without manual correction, achieved a lower test AUC of 0.708 with 71.87% accuracy. Case 2 with manually corrected alignment achieved AUC of 0.783, underscoring that precise tumor segmentation is critical for reliable radiomic feature extraction and model performance.
Comparable to prior published studies. A summary of six prior studies on radiomic PD-L1 prediction in NSCLC shows AUC values ranging from 0.71 to 0.829. The AUC of 0.783 in the current study falls within this range, confirming that the approach is feasible and the results are competitive despite the smaller sample size of 105 patients compared to datasets of 255-399 patients in some prior work.
PET features contributed uniquely in this study. Unlike one prior large-dataset study (Jiang et al., n=399) which found CT features more predictive than PET features, this study found that PET-derived features made important independent contributions in the selected model. The authors attribute the discrepancy to the prior study's use of lower-resolution PET images and suboptimal CT segmentation methodology.
Alignment with other multi-modal studies. Studies by Li et al. and Zhao et al. that combined clinical, PET, and CT features into joint models found that fusing PET and CT information improves PD-L1 prediction, consistent with this study's finding that the best-performing models used features from both modalities. The convergent evidence across multiple research groups strengthens the case for multimodal radiomic analysis.
The role of the tumor microenvironment. Recent research shows that PD-L1 blockade achieves antitumor efficacy primarily when CD8+ tumor-infiltrating lymphocytes (TILs) are present. Tumors with high PD-L1 expression and CD8+ TILs (Tumor Microenvironment Type I) respond best to anti-PD-1/PD-L1 therapy. Extending radiomic models to capture immune microenvironment characteristics alongside PD-L1 expression could further refine patient selection.
A proof-of-principle for non-invasive PD-L1 assessment. This study demonstrates that radiomic features from 18F-FDG-PET/CT images -- scans already performed routinely for NSCLC staging -- can meaningfully predict PD-L1 expression status. For patients where biopsy is technically impossible, high-risk, or impractical, this approach could substitute for or triage tissue-based PD-L1 testing.
High sensitivity prioritizes patient benefit. The 90% sensitivity of the best model ensures that most PD-L1 positive patients are correctly identified, minimizing the risk of inappropriately denying immunotherapy to eligible patients. The lower specificity (66.67%) means some PD-L1 negative patients may be classified as positive, which carries less clinical risk than missing true positives who would benefit from ICI treatment.
Limitations acknowledged. The study is limited by its single-center design, small sample size, and reliance on images from a single PET/CT scanner. Manual segmentation is time-consuming and poorly reproducible, particularly for tumors with spiculated margins. The ground truth PD-L1 testing used different antibody clones in some cases, which can affect comparability.
Future research agenda. Planned directions include expanding the patient cohort with multi-center data from different scanners, testing automated segmentation methods, predicting quantitative PD-L1 expression levels rather than binary status, and incorporating deep learning features to complement conventional radiomic analysis. These steps would move the technology closer to routine clinical implementation.