Lung cancer is the leading cause of cancer death worldwide, and over 70% of NSCLC cases are diagnosed at advanced stages with poor prognoses. Accurate prognostication - predicting how a patient's disease will progress - is essential for making informed, personalized treatment decisions and helping patients understand their disease trajectory.
PET/CT imaging using the radiotracer FDG is routinely performed for NSCLC patients as part of standard disease workup. FDG (fluorodeoxyglucose) is a glucose analog that accumulates preferentially in metabolically active cancer cells, making PET/CT scans a rich source of information about tumor biology that could be harnessed for patient risk stratification.
Radiomics - the extraction of quantitative features from medical images - offers a non-invasive window into tumor biology. However, a major limitation of most radiomic approaches is that the extracted features are identified through statistical data mining rather than being grounded in a biological understanding of how cancer works. This makes them difficult to interpret and reduces confidence in their clinical validity.
Tumors are not uniform inside - they contain populations of cells with different behaviors and metabolic activities. Understanding the spatial organization of these subpopulations within a tumor could reveal important information about how aggressive the cancer is. This study introduced a new imaging biomarker designed to capture exactly this kind of spatial information.
The normalized hotspot-to-centroid distance (NHOC) is a novel biomarker that measures how far the most metabolically active part of a tumor is from the tumor's geometric center. On PET scans, the metabolic hotspot corresponds to the area with the highest FDG uptake (SUVmax), reflecting the region of greatest glucose consumption. The centroid is the geometric center of the tumor volume.
NHOC is grounded in a cancer evolutionary model. As cancers become more aggressive, the most active cancer cell subpopulations are theorized to drift toward the periphery of the tumor as competition between cell subpopulations intensifies. A hotspot located closer to the tumor's edge suggests more aggressive evolutionary dynamics. NHOC captures this relationship by expressing the distance as a ratio normalized by the tumor's metabolic spherical radius, making it independent of tumor size.
Higher NHOC values indicate a more peripheral metabolic hotspot and are predicted to correlate with more aggressive disease. Because the measure is dimensionless and accounts for tumor size, it can be compared meaningfully across patients with tumors of different sizes. This mathematical grounding in cancer biology gives NHOC an advantage in explainability over features identified purely through statistical analysis.
The Warburg effect provides the biological basis for why FDG-PET can detect these aggressive cell populations. Cancer cells with high metabolic activity preferentially consume glucose even in the presence of oxygen - a phenomenon known as the Warburg effect. This causes aggressive tumor regions to stand out clearly on FDG-PET imaging, making the spatial location of these hotspots measurable and clinically meaningful.
This retrospective multi-center study included 285 NSCLC patients from Imperial College Healthcare NHS Trust as the discovery cohort, with external validation across four additional UK cancer centers. The external centers - King's College London, Royal Marsden Hospital, Mount Vernon Hospital, and Nottingham University Hospital - contributed a combined 215 patients, enabling rigorous testing of whether the biomarker performs consistently across different institutions, scanner types, and patient populations.
All patients included had histologically confirmed NSCLC, a tumor volume of at least 5 ml, and a pre-treatment FDG-PET/CT scan available. Patients receiving only palliative treatment or surgery were excluded, and the analysis focused on patients undergoing radical radiotherapy with or without chemotherapy. The key outcome was 3-year overall survival.
Multi-region segmentation captured not just the primary tumor but also the surrounding tissue environment. Three volumes of interest were defined for each patient: the primary tumor lesion, a peri-tumoral annular shell extending 1 cm outward from the tumor boundary, and a region of normal-appearing background lung parenchyma. This spatial approach was designed to capture tumor-microenvironment interactions that may reflect underlying cancer biology.
Inter-scanner variability was addressed through image harmonization techniques. Because patients were scanned on different PET/CT systems across institutions, ComBat statistical harmonization was applied after radiomic feature extraction to reduce batch effects. Only features with high reproducibility between two expert readers - measured by intraclass correlation coefficient above 0.8 - were retained for analysis.
From 3,996 radiomic features initially extracted per patient, a rigorous filtering process reduced this to 87 features with both high reproducibility and statistical significance for survival prediction. Features with redundant information (high pairwise correlation) were removed, and only those passing univariable Cox regression at a false discovery rate below 1% were retained. This disciplined reduction was essential for preventing overfitting and building a model that would generalize.
A radiomics predictive vector (RPV) was developed using elastic net regularization and Cox regression, comprising just 9 radiomic features. These included texture features from the peri-tumoral annulus on CT, wavelet and fractal features from the tumor core on CT, and wavelet-texture features from the peri-tumoral annulus on FDG-PET. The fact that peri-tumoral features dominated the signature supported the hypothesis that peripheral tumor regions carry important prognostic information.
The composite biomarker nLCEV (non-invasive lung cancer evolution vector) was formed by combining NHOC, the RPV, and disease stage. Disease stage was included because it retained independent significance in multivariable Cox analysis. By integrating a biologically-motivated spatial metric (NHOC) with radiomics-derived texture information and standard staging, nLCEV was designed to capture complementary dimensions of disease biology in a single prognostic score.
Patients were classified into high-risk and low-risk groups using k-means clustering based on nLCEV scores. This unsupervised grouping allowed assessment of whether the composite score could meaningfully separate patients with very different survival probabilities without requiring a manually selected cutoff, improving the practical applicability of the tool across different clinical populations.
Both NHOC and the radiomics predictive vector RPV demonstrated statistically significant independent prognostic value. NHOC alone achieved a hazard ratio of 2.52, meaning patients with high NHOC had more than twice the risk of death within 3 years compared to those with low NHOC. RPV achieved an even stronger hazard ratio of 2.68. Importantly, neither metric was simply a proxy for tumor size - their correlations with tumor volume were moderate rather than strong.
The composite biomarker nLCEV achieved an area under the ROC curve of 0.76 for 3-year overall survival prediction in the internal validation set. This outperformed NHOC alone (AUC 0.68), RPV alone (AUC 0.72), and the conventional metric SUVmax (AUC 0.66). An AUC of 0.76 means the model correctly ranked the survival risk of randomly selected patient pairs about 76% of the time.
Critically, nLCEV achieved statistically significant patient stratification in all four independent external validation cohorts. Hazard ratios ranged from 2.21 at the Royal Marsden to 4.14 at Nottingham University Hospital, with all cohorts showing p-values below 0.05. This consistency across institutions with different scanners, patient demographics, and staging distributions is the strongest evidence for the biomarker's robustness.
SUVmax and the conventional disease stage-metabolic model (SMM) failed to achieve statistically significant stratification in some external cohorts. SUVmax did not reach significance at the Royal Marsden or Nottingham hospitals, and SMM failed in all four external cohorts despite performing well in the discovery set - a pattern indicative of overfitting to the training data. nLCEV was the only biomarker to perform consistently across all cohorts.
The dominance of peri-tumoral and peripheral features in the prognostic signature provides biological support for the NHOC hypothesis. The most heavily weighted radiomic feature in RPV was a wavelet-texture descriptor from the peri-tumoral annulus on FDG-PET. Features from the peri-tumoral region carried positive predictive weights for patient mortality, while features from the tumor core carried negative weights - consistent with the idea that peripheral tumor activity drives worse prognosis.
This biologically grounded approach to feature design is what sets nLCEV apart from most radiomic signatures. Conventional radiomics relies on post-hoc interpretation tools like SHAP values or saliency maps to explain model predictions, but these methods reveal correlations rather than mechanisms. By building NHOC from a cancer evolutionary model first, the study embeds biological justification directly into the feature construction stage, enhancing transparency and clinical trustworthiness.
NHOC should currently be regarded as an imaging-based surrogate of spatial heterogeneity rather than a directly validated measure of clonal cell competition. While the metric is motivated by evolutionary cancer biology - specifically the theory that aggressive subclones drift to the tumor periphery as they compete for resources - direct visualization of clonal dynamics would require additional specialized imaging tracers beyond standard FDG-PET.
The nLCEV framework also offers practical advantages for clinical implementation. A semi-automated segmentation software was developed as part of this work to assist in computing NHOC and extracting peri-tumoral radiomic features from PET/CT scans, reducing the manual burden on radiologists and improving reproducibility for potential future clinical deployment.
A key limitation is that all imaging metrics were derived from the primary tumor only, not accounting for metastatic disease burden. In patients with distant spread, the biology of metastatic lesions - their number, size, and metabolic activity - may carry independent prognostic significance. Advances in AI-based whole-body PET analysis now allow automated measurement of total tumor volume and metabolic activity across all sites simultaneously, and future extensions of this framework could incorporate these whole-body metrics.
The study excluded patients with low FDG uptake, which biases the cohort toward more aggressive tumors. Lung adenocarcinomas with low metabolic activity often have better prognoses, and they were excluded because reliable segmentation on PET is not feasible at low uptake levels. This limits the applicability of the current biomarker to the FDG-avid subset of NSCLC.
The retrospective design and relatively small external validation cohorts (38-63 patients each) mean prospective validation is still needed. Despite the biomarker performing consistently across all five cohorts, larger prospective studies across more diverse international populations would be necessary before clinical adoption. Differences in tumor staging and histological composition between cohorts also contributed to variability in the magnitude of prognostic effects observed.
Future work could integrate histological and molecular measures of intra-tumoral heterogeneity to enhance nLCEV further. A pilot study included in the supplementary materials showed promising results when histological tumor features were combined with the imaging-based nLCEV. Incorporating spatial transcriptomics, sarcopenia measures, and treatment response delta-radiomics could strengthen the biological and prognostic depth of the composite biomarker.
NHOC is the first imaging biomarker for NSCLC prognosis to be derived from a cancer evolutionary model, providing biological justification that most radiomic features lack. By measuring the relative peripherality of a tumor's metabolic hotspot on FDG-PET, it captures a meaningful dimension of tumor spatial organization that correlates with patient survival.
When combined with peri-tumoral radiomic texture features and disease stage in the nLCEV composite score, prediction accuracy improves substantially over any individual component. The final model outperformed the widely used SUVmax metric and a conventional multi-variable metabolic staging model in both accuracy and generalizability across independent cohorts.
The validated external performance of nLCEV across four hospitals with different scanners, patient demographics, and imaging protocols makes it one of the more rigorously tested NSCLC radiomic prognostic signatures published to date. This consistent multi-institution performance is a prerequisite for any imaging biomarker moving toward clinical deployment.
The non-invasive nature of this approach is particularly valuable in clinical scenarios where tissue sampling is difficult or yields equivocal results. Deriving prognostic information entirely from routine pre-treatment PET/CT scans means the biomarker adds value without additional procedures, aligning well with the growing emphasis on personalized treatment planning in NSCLC.