Noninvasive Prediction of Programmed Cell Death Protein-Ligand 1 Expression in Locally Advanced Non-small Cell Lung Cancer by 18F-Fluorodeoxyglucose Positron Emission Tomography/Computed Tomography-Based Metabolic Habitats: A Multicenter Radiomic and Biological Study

Ann Surg Oncol 2025 AI 7 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
The Case for Noninvasive PD-L1 Prediction in LA-NSCLC

PD-L1 expression guides immunotherapy but is difficult to assess reliably. Locally advanced non-small cell lung cancer, representing stage III disease, accounts for approximately 20 to 25% of NSCLC cases and carried a 5-year survival rate of only 10 to 30% in the pre-immunotherapy era. Immune checkpoint inhibitors have substantially improved outcomes, but their use depends on PD-L1 expression measured as a tumor proportion score -- the only predictive biomarker currently available in clinical practice.

Tissue biopsy-based PD-L1 testing has significant practical limitations: sampling bias arises because a single biopsy captures only a small, potentially unrepresentative tumor region; results are poorly reproducible due to the heterogeneous nature of PD-L1 expression across tumor sites; the procedure is invasive; assays are not rapid; and costs are considerable. In some cases, tissue quantity or quality is insufficient to yield actionable results. Noninvasive longitudinal approaches are urgently needed to complement or partially replace tissue-based PD-L1 assessment.

18F-FDG PET/CT is a routine noninvasive molecular imaging method that reveals glycolysis status throughout the tumor and body. PD-L1 expression has a bidirectional positive feedback relationship with glycolysis: PD-L1 upregulates glycolysis in tumor cells, and glycolytic pathways in turn promote PD-L1 expression. This mechanistic link positions 18F-FDG PET/CT as a logical imaging platform for predicting PD-L1 status noninvasively.

TL;DR: PD-L1 tissue testing is limited by sampling bias and invasiveness, while a bidirectional PD-L1-glycolysis feedback relationship makes FDG PET/CT a biologically motivated noninvasive alternative for predicting PD-L1 status in LA-NSCLC.
Pages 2-3
From Whole-Tumor Radiomics to Metabolic Habitat Analysis

Standard radiomics ignores intratumoral spatial heterogeneity. Prior studies have demonstrated that 18F-FDG PET/CT radiomics predicts PD-L1 expression in NSCLC with AUC values ranging from 0.604 to 0.970. However, these studies extracted features from the entire tumor region, implicitly assuming homogeneous distribution of biological properties within the tumor. This assumption ignores density and metabolic differences caused by intratumoral heterogeneity, resulting in loss of spatial imaging information and poor interpretability of the resulting models.

Habitat analysis addresses this limitation by capturing spatial heterogeneity within tumors. Metabolic habitat images derived from 18F-FDG PET/CT uncover subregional structural features, reflect the degree of metabolism within each habitat, and quantify metabolic differences between habitats. Prior work has used habitat images from PET/CT to identify benign and malignant lesions, classify pathologic subtypes, evaluate immunotherapy response, and predict recurrence in NSCLC, consistently showing that metabolic habitats more effectively capture microenvironmental changes than conventional whole-tumor radiomics.

No prior study had systematically elucidated the association between different metabolic habitat radiomics and PD-L1 expression in LA-NSCLC. This study filled that gap by constructing imaging biomarkers for predicting immune status using habitat radiomic features from four distinct metabolic subregions, and by exploring the biological characteristics of NSCLC through analysis of glycometabolism-related gene expression in relation to PD-L1.

TL;DR: Conventional radiomics treats the tumor as uniform and loses spatial heterogeneity information, motivating habitat analysis that partitions the tumor into metabolically distinct subregions and models each independently.
Pages 2-5
Study Design: Habitat Generation and Model Construction

219 patients from two centers, four metabolic habitats, Extra Trees classifier. The study retrospectively enrolled 219 biopsy-confirmed stage III NSCLC patients from two Chinese medical centers who received baseline 18F-FDG PET/CT before treatment and had PD-L1 expression testing. The training cohort (175 patients from Shandong Cancer Hospital) and external testing cohort (44 patients from Second Qilu Hospital of Shandong University) had 97 and 25 PD-L1-positive cases, respectively. An additional 1,043 TCGA patients provided transcriptomic data for gene analysis.

A five-stage imaging preprocessing pipeline addressed scanner heterogeneity: intensity normalization, spatial PET-to-CT registration, resampling to uniform resolution, super-resolution reconstruction to recover PET image detail, and image discretization. Super-resolution reconstruction was applied because prior studies have shown that low PET resolution causes loss of image information that degrades model performance. Tumor segmentation was performed manually using 3D Slicer by experienced radiologists.

Habitat segmentation used the Otsu binary classification algorithm, applied separately to PET and CT images to classify each into high and low grayscale subregions based on maximum interclass variance of voxel intensities. Intersection of the PET and CT binary segmentations yielded four distinct metabolic habitat subregions: high-glycolytic/high-density (Habitat 1), low-glycolytic/low-density (Habitat 2), low-glycolytic/high-density (Habitat 3), and high-glycolytic/low-density (Habitat 4).

From each habitat and from the whole tumor, 1,015 radiomic features were extracted per region including intensity, shape, texture, and wavelet features. A three-step feature selection procedure applied a Mann-Whitney U test, Spearman correlation filtering to remove redundant features, and LASSO with tenfold cross-validation to select features with non-zero coefficients. The Extra Trees classifier was selected as the prototype model for its robustness, variance reduction, and computational efficiency, and the final metabolic habitat model was compared against a whole-tumor radiomics model and a clinical model.

TL;DR: Four metabolic habitats were generated from FDG PET/CT using Otsu thresholding on PET and CT images, and Extra Trees classifiers trained on LASSO-selected radiomic features from each habitat predicted PD-L1 expression in 219 patients from two centers.
Pages 6-8
Model Performance and Habitat Comparison

Metabolic habitat model achieves AUC 0.833 and outperforms all comparators. The metabolic habitat model achieved an AUC of 0.833 (95% CI 0.775 to 0.892) in the training cohort and 0.786 (95% CI 0.649 to 0.923) in the external testing cohort, with sensitivity of 80.0% and specificity of 63.2% on the test set. This performance exceeded the whole-tumor radiomics model (AUC 0.806 training, 0.639 testing) and the clinical model (AUC 0.621 training, 0.581 testing). Decision curve analysis confirmed that the metabolic habitat model generated greater clinical net benefit across a broad range of threshold probabilities.

Metabolic parameter analysis across the four habitats revealed consistent patterns. Habitat 1 (high-glycolytic/high-density) exhibited the highest SUVmax and total lesion glycolysis values, indicating densely cellular regions with maximally active glycolysis and heightened energy demands consistent with active tumor proliferation or immune evasion. Habitats 1 and 4 showed the highest SUVmean, both significantly exceeding Habitats 2 and 3. Metabolic tumor volume was highest in Habitats 1 and 3, reflecting their larger spatial extent within tumors.

Spatial characteristic analysis demonstrated that only Habitat 1 (high-glycolytic/high-density) showed statistically significant positive correlation with PD-L1 expression. Both the number of Habitat 1 voxels (OR 1.014 per 1,000 voxels, p = 0.042) and the volume fraction of Habitat 1 (OR 8.84, p = 0.043) were independent predictors of PD-L1 positivity. The other three habitats showed no significant correlation, confirming the specificity of Habitat 1 for immune expression prediction. The larger the volume fraction of high-glycolytic/high-density tissue within the tumor, the greater the probability of PD-L1 positivity.

TL;DR: The metabolic habitat model achieved AUC 0.833/0.786 in training and testing, outperforming whole-tumor radiomics (0.639 testing) and clinical models, with Habitat 1 (high-glycolytic/high-density) as the sole spatial predictor significantly associated with PD-L1 expression.
Pages 8, 10, 11
Biological Validation: Glycolysis Genes and PD-L1 Crosstalk

Four glycolysis-related genes link tumor metabolism to immune escape. Transcriptomic analysis of 1,043 TCGA NSCLC patients (541 LUAD, 502 LUSC) identified 62 differentially expressed genes between PD-L1-high and PD-L1-low tumors. Intersection with 753 glycolysis-related genes revealed four genes at the convergence of glycolytic metabolism and PD-L1 regulation: IFNG, IL2RA, HK3, and MYCN.

PD-L1 expression was positively correlated with IL2RA (R = 0.522, p less than 0.001), IFNG (R = 0.468, p less than 0.001), and HK3 (R = 0.458, p less than 0.001). Each of these three genes has established roles in tumor proliferation, metastasis, antigen presentation, and T-cell regulation that promote PD-L1 expression: IL2RA regulates the classical Akt/mTOR pathway, IFNG is a key T-cell-secreted cytokine that upregulates PD-L1 as an immune evasion response, and HK3 encodes hexokinase-3, which mediates early glycolysis.

MYCN was negatively correlated with PD-L1 expression (R = -0.201, p less than 0.001), a finding attributed to MYCN's activation of multiple competing energy metabolic pathways including aerobic glycolysis, the tricarboxylic acid cycle, oxidative phosphorylation, and glutaminolysis. These contradictory energy metabolic effects may limit net glycolytic flux in ways that reduce the glycolysis-PD-L1 positive feedback loop. This gene-level biological validation establishes a molecular basis for why high-glycolytic imaging habitats correspond to elevated PD-L1 expression.

TL;DR: Four glycolysis-PD-L1 genes were identified in TCGA data: IFNG, IL2RA, and HK3 are positively correlated with PD-L1 expression while MYCN is negatively correlated, providing molecular biological validation of the PET/CT habitat-PD-L1 imaging findings.
Pages 11-12
Clinical and Methodological Advantages of Habitat Mapping

Actionable biopsy guidance and a noninvasive alternative for immunotherapy decisions. The metabolic habitat model offers two categories of clinical benefit. First, as a noninvasive PD-L1 prediction tool, it can reduce unnecessary biopsies in patients with high predictive confidence and provides an alternative for patients ineligible for invasive procedures, enabling immunotherapy decisions from routine imaging without additional cost. Second, the identification of PD-L1-associated metabolic habitats provides actionable biopsy targeting guidance: sampling high-glycolytic/high-density regions while avoiding metabolically reduced areas optimizes tissue sampling accuracy and reduces false-negative rates in heterogeneous LA-NSCLC.

Methodologically, habitat radiomics represents an upgrade from conventional whole-tumor radiomics by precisely quantifying the morphological and functional differences caused by intratumoral heterogeneity. The use of the Otsu algorithm for habitat segmentation provides better interpretability and repeatability than other segmentation approaches. Super-resolution reconstruction was essential for preserving PET image information, as studies have shown that low PET resolution is a key cause of model performance degradation in radiomics analyses combining PET and CT data.

The bidirectional glycolysis-PD-L1 crosstalk identified in both imaging and genomic analyses creates a coherent translational framework: FDG PET/CT habitats reflect the glycometabolic microenvironment, glycolytic genes regulate PD-L1 expression, and PD-L1 in turn promotes further glycolysis. This longitudinal crosstalk between genes, metabolism, and molecular imaging provides a theoretical foundation for future metabolic immunotherapy approaches that modulate glycolysis pathways to influence immune checkpoint expression.

TL;DR: The habitat model enables noninvasive PD-L1 assessment and provides actionable biopsy targeting guidance, while the metabolic-genomic crosstalk framework supports future strategies to modulate glycolysis as a route to altering PD-L1 expression.
Page 12
Limitations and Future Directions

Promising results constrained by retrospective design and testing cohort size. The retrospective study design introduces the possibility of clinical data loss and selection bias, necessitating large-scale prospective studies for model validation. The external testing cohort of 44 patients successfully validated model performance but is too small to extrapolate findings to broader populations: small testing sets may overestimate predictive efficacy due to reduced statistical power and susceptibility to sampling bias. Multi-institutional prospective studies with larger cohorts are needed to confirm generalizability.

Some PD-L1 expression data were derived from puncture biopsy pathology rather than surgical specimens. While prior studies have shown high consistency between puncture and surgical samples, spatial heterogeneity in NSCLC may introduce sampling bias particularly for tumors with low PD-L1 expression. Future studies should incorporate multiregional biopsy or liquid biopsy-based PD-L1 profiling to better capture intratumoral heterogeneity and reduce this limitation.

The metabolic habitat model demonstrated biological value in relating imaging subregions to PD-L1 expression, but further pathological, genetic, and mechanistic studies are needed to confirm accuracy of the biological interpretations. Future work should directly investigate the relationship between metabolic habitats and immunotherapy response outcomes and patient prognosis, extending the model from a PD-L1 prediction tool to a comprehensive predictor of treatment benefit and disease trajectory.

TL;DR: The habitat radiomics model requires prospective multi-institutional validation with larger cohorts, liquid biopsy-based PD-L1 reference standards, and direct correlation with immunotherapy response outcomes to confirm clinical readiness.
Citation: Open Access, 2025. Available at: PMC12589310.