The burden of lung cancer. Non-small cell lung cancer (NSCLC) is the most common type of lung cancer and accounts for roughly one-quarter of all cancer-related deaths worldwide. About 85% of lung cancers are NSCLC, and most are adenocarcinomas, a specific cell type that has become the focus of targeted treatment strategies.
Targeted therapies change the game. Modern lung cancer care increasingly relies on identifying specific genetic mutations - particularly in genes called EGFR and KRAS - to guide treatment. Drugs that target these mutations can dramatically improve survival compared to conventional chemotherapy, making mutation testing a critical first step in patient management.
Access to testing remains uneven. The gold standard for detecting these mutations is next-generation sequencing (NGS), a sophisticated laboratory process that reads the tumor's genetic code. However, NGS requires adequate tissue samples, specialized equipment, trained staff, and significant cost - barriers that prevent uniform access across all medical centers.
The need for noninvasive alternatives. When biopsies cannot be obtained - due to tumor location, patient frailty, or insufficient tissue - clinicians are left without the molecular information needed to select the best treatment. This gap motivates the search for noninvasive approaches that could reveal genetic information from standard imaging scans already performed in routine care.
Radiomics defined. Radiomics is a computational approach that extracts hundreds of quantitative features from medical images - features invisible to the naked eye - to reveal hidden information about tumor biology. Rather than replacing imaging, radiomics uses images that patients already undergo, making it inherently noninvasive and cost-efficient.
PET/CT as the imaging foundation. This study focused on [18F]FDG PET/CT scans, which measure how much sugar (glucose) a tumor consumes - a proxy for how metabolically active and aggressive it is. Cancer cells typically consume far more glucose than normal tissue, a phenomenon that lights up on PET scans and can reflect the tumor's molecular characteristics.
Capturing tumor heterogeneity. Radiomic features can quantify aspects like texture, shape, and intensity patterns within a tumor. These patterns may reflect underlying biological differences - such as whether certain genes are mutated - that alter how cancer cells behave metabolically.
The radiogenomics concept. The field of radiogenomics explores the relationship between imaging features and genomic characteristics. If specific radiomic patterns reliably correlate with specific mutations, a scan could serve as a surrogate genetic test - a concept sometimes called a "digital biopsy."
Retrospective cohort at a single center. The study enrolled 105 patients with histologically confirmed NSCLC who had both baseline [18F]FDG PET/CT scans and NGS results from lung tissue. All scans were performed between January 2023 and December 2024 at Fondazione IRCCS San Gerardo dei Tintori in Monza, Italy.
Strict inclusion criteria. Patients needed to have a confirmed NSCLC diagnosis, a PET/CT scan performed within one month of diagnosis, and molecular profiling from lung lesions specifically. An initial pool of 140 patients was reviewed; 35 were excluded because their tumors could not be accurately segmented on PET images due to overlap with nearby lymph nodes or very low glucose uptake.
Patient characteristics. The final group of 105 patients had an average age of 70 years and was 46% female. Most tumors (87%) were adenocarcinomas. NGS revealed KRAS mutations in 44% of patients, EGFR mutations in 13%, and other mutations (including MET, ALK, and BRAF) in 20% of patients.
Two independent scanner datasets. A key methodological strength was the use of two different PET/CT scanners - the Discovery MI (55 patients) and the Discovery IQ (50 patients) - which allowed researchers to identify radiomic features in one group and validate them independently in the other, a more rigorous approach than single-dataset analysis.
Comprehensive feature extraction. Using the open-source Pyradiomics tool - which follows international image biomarker standardization (IBSI) guidelines - the researchers extracted 766 distinct quantitative features from each tumor's PET scan region. These included basic shape measurements, first-order intensity statistics, and more complex higher-order texture features.
Multiple image processing approaches. To capture different aspects of tumor biology, features were computed not just on the original PET image but also on versions processed through wavelet transforms and Laplacian of Gaussian (LoG) filtering at different scales. Each processing approach reveals different patterns within the tumor's metabolic signal.
Rigorous feature selection to prevent overfitting. Given the small sample size relative to the large number of features, the researchers applied a multi-step selection process. Features were tested 500 times using random subsampling combined with LASSO regularization - a statistical technique that aggressively eliminates redundant or weakly predictive features. Only features appearing in more than 35% of the 500 models were retained.
Independent validation step. Features identified as promising in the Discovery MI dataset were then tested on the entirely separate Discovery IQ dataset, using metrics including area under the curve (AUC), statistical significance testing, and odds ratios. This cross-scanner validation is essential to demonstrate that a feature reflects true biology rather than scanner-specific noise.
No EGFR signal found. Despite the large number of features evaluated, no radiomic feature showed a significant correlation with EGFR mutation status. This null result for EGFR may partly reflect the small number of EGFR-positive patients in the cohort (only 14 of 105), which limits statistical power to detect true associations.
A promising KRAS finding. For KRAS mutations, eight features initially showed correlation in the Discovery MI dataset. When tested on the independent Discovery IQ dataset, one feature - FBS_glcm_MCC - retained statistical significance. This feature achieved an AUC of 0.68 (where 0.5 is chance and 1.0 is perfect) with a p-value of 0.04 and an odds ratio of 0.65.
What FBS_glcm_MCC measures. This feature is the Maximal Correlation Coefficient derived from the Gray-Level Co-occurrence Matrix, applied after standardized intensity binning. It quantifies texture complexity: lower values indicate more irregular, heterogeneous uptake patterns within the tumor, while higher values reflect more uniform glucose distribution. KRAS-mutated tumors showed lower values - meaning more complex, irregular metabolic texture.
Biological interpretation. KRAS mutations are known to rewire how cancer cells use energy, increasing glycolysis, glutamine use, and fatty acid synthesis. These metabolic changes may produce spatially irregular glucose uptake patterns that manifest as lower texture uniformity on PET scans - essentially making KRAS tumors metabolically messier in a way that imaging can detect.
Consistent with the literature for KRAS. A prior study by Zhang et al. developed a PET-based KRAS prediction model using three radiomic features plus smoking status, achieving AUCs of 0.73 to 0.75. Although this study achieved a lower AUC of 0.68 with a single feature, the core finding - that heterogeneity-related PET texture features correlate with KRAS mutation - aligns well across both studies.
Conflicting with prior EGFR findings. Several earlier studies reported significant associations between PET/CT radiomic features and EGFR mutation status, with some achieving AUCs above 0.80. The failure to replicate these findings here likely reflects the small EGFR-positive sample and the fact that EGFR-mutated tumors tend to have lower, more uniform glucose uptake - making them harder to segment and potentially excluded from this cohort.
Clinical context of KRAS detection. KRAS mutations have historically been considered untreatable, but recent approvals of KRAS-targeting drugs have changed this landscape. Early and noninvasive detection of KRAS mutations could help select patients who may benefit from these newer therapies and guide decisions about immunotherapy, which tends to be less effective in KRAS-mutated NSCLC.
Current performance is insufficient for standalone use. An AUC of 0.68 is too low to replace molecular testing in clinical practice. The researchers are clear that these results are exploratory and should not yet change clinical workflows. The value lies in demonstrating proof of concept and motivating larger prospective studies with standardized methods.
Small sample size. With only 105 patients - and fewer for subgroup analyses by mutation type - the study is underpowered for definitive conclusions. The risk of overfitting is real, even with the repeated cross-validation and LASSO regularization applied. Larger cohorts are needed to confirm these findings.
Scanner variability and harmonization challenges. Using two different PET/CT scanner models is both a strength (enabling cross-validation) and a limitation. Radiomic features can vary between scanners even when imaging the same object, and the small sample size precluded formal harmonization methods that are typically applied in multi-center radiomics studies.
Segmentation exclusions may bias results. 35 patients were excluded because tumors could not be cleanly delineated on PET images. This non-random exclusion - particularly affecting tumors with low glucose uptake or complex anatomy - may have disproportionately removed EGFR-positive cases, since EGFR tumors typically show lower PET signal.
Static imaging only. This study used standard static PET acquisitions rather than dynamic imaging that tracks tracer uptake over time. Dynamic PET can provide richer information about tumor physiology and might improve the predictive power of radiomic models. Integration with CT and MRI features could also add complementary structural information.
A step toward noninvasive molecular profiling. This study adds to a growing body of evidence that PET/CT radiomics can capture biologically meaningful information about tumor genetics - specifically, that texture complexity in FDG uptake may serve as a surrogate for KRAS mutation status in NSCLC patients.
Practical implications if validated. Patients with NSCLC already undergo PET/CT as standard of care. If radiomic analysis of these scans could reliably identify mutation status, it would provide an additional layer of molecular information without extra cost, invasiveness, or delay. This could be especially valuable when tissue is limited or NGS is not accessible.
The path forward requires larger studies. The researchers call for prospective, multicenter trials with standardized imaging protocols, harmonized feature extraction methods, and larger patient cohorts. These studies would need to demonstrate not just statistical associations but clinical utility - meaning that radiomic predictions actually improve patient outcomes.
Integration with multimodal data. Future models combining PET radiomics with CT texture, clinical features such as smoking history, liquid biopsy results, and even artificial intelligence-driven pattern recognition could substantially improve predictive performance. The current work lays a foundation for that more comprehensive approach.