Non-small cell lung cancer (NSCLC) is not one disease but a collection of molecularly distinct subtypes, each driven by specific genetic mutations that activate different growth pathways, attract different patterns of spread, and respond to different targeted therapies. Identifying the specific oncogenic driver mutation guiding a patient's tumor is now central to modern lung cancer treatment planning.
Molecular testing through tissue biopsy remains the definitive method for driver identification, but biopsy is invasive, sometimes technically impossible (particularly in centrally located or small tumors), and may not capture the full genetic heterogeneity of a tumor that has spread to multiple sites. These limitations motivate the search for non-invasive predictors of molecular subtype.
Radiogenomics is the emerging field that links quantitative and qualitative CT imaging features to underlying tumor genomics. The hypothesis is biologically grounded: different oncogenic driver mutations alter how cancer cells grow, invade, and interact with surrounding tissue, creating characteristic patterns of tumor shape, density, location, and spread that are visible on CT scans and potentially detectable by both radiologists and AI-based image analysis tools.
This comprehensive review covers nine major targetable oncogenic drivers in NSCLC: EGFR (epidermal growth factor receptor), ALK (anaplastic lymphoma kinase), KRAS (Kirsten rat sarcoma viral oncogene), ROS1 (ROS proto-oncogene 1), RET (rearranged during transfection), MET exon 14, BRAF, HER2 (human epidermal growth factor receptor 2), and NTRK (neurotrophic tropomyosin receptor kinase). Together these mutations account for a large fraction of actionable NSCLC cases.
For each driver, the review synthesizes three types of imaging data: conventional radiological features visible to expert radiologists (tumor location, size, density, margins, and presence of ground-glass opacity), radiomic features extracted computationally from pixel-level image analysis, and metastatic patterns describing where each driver preferentially spreads when the disease becomes advanced.
Understanding metastatic patterns by driver has direct clinical value beyond diagnosis. Knowing that EGFR-mutant tumors favor brain and miliary lung metastases while ALK-rearranged tumors favor lymphangitic spread helps clinicians select appropriate staging and surveillance protocols - for example, choosing whether to include brain MRI in the initial workup or monitor closely for leptomeningeal disease in patients already known to carry specific mutations.
EGFR mutations are the most common targetable driver in Western populations, especially in non-smoking women, and carry a characteristic CT profile: smaller tumors, more peripheral location, frequent ground-glass opacity (GGO) or mixed solid-GGO pattern (part-solid), air bronchograms (visible air-filled airways within the tumor), and spiculated margins. These features reflect the adenocarcinoma histology that almost exclusively harbors EGFR mutations.
EGFR's metastatic signature is distinctive: miliary lung metastasis (a pattern of innumerable tiny nodules spread throughout both lungs) and brain metastases are the dominant distant spread patterns, and leptomeningeal metastasis (spread along the brain's lining) is particularly associated with EGFR exon 20 insertions and T790M resistance mutations. This brain tropism is a major clinical challenge, as standard systemic therapies penetrate the blood-brain barrier poorly.
ALK rearrangements present with a contrasting CT profile: solid tumors without significant GGO, more central or perihilar location, frequent mucus plugging of airways, and a tendency toward very bulky mediastinal lymphadenopathy out of proportion to the primary tumor size. ALK tumors are more likely to have pericardial effusion and large pleural effusions. Metastatic spread in ALK-rearranged NSCLC favors lymphangitic carcinomatosis (spread along lymphatic channels within the lung) and unusual metastatic sites including the pericardium.
KRAS mutations - common in heavy smokers - produce a CT profile that reflects their epidemiology and biology: solid tumors without GGO (pure solid), round or lobulated margins, larger size at presentation, and location in emphysematous lung parenchyma with background smoking-related damage. KRAS tumors are more likely to have cavitation (air pockets within the tumor mass). Metastatic spread favors lungs and brain, with pleural spread being less characteristic than in EGFR or ALK-mutant disease.
ROS1 fusions, which respond to similar targeted drugs as ALK rearrangements, tend to present as peripheral solid tumors with spiculated margins and frequent involvement of both the primary tumor site and extensive regional lymph nodes. ROS1 is notable for its tendency to spread to unusual distant nodal stations, including supraclavicular and infradiaphragmatic lymph nodes that are rarely involved in other NSCLC subtypes. This extensive nodal involvement can mimic lymphoma on imaging.
RET fusions are associated with a radiological profile that includes peripheral, predominantly solid nodules often with spiculation, and a particularly high burden of distant metastases at presentation. Brain and bone metastases occur at high rates, and lymphangitic spread is common. The aggressive metastatic phenotype of RET-rearranged NSCLC at diagnosis helps explain why targeted therapy with specific RET inhibitors significantly improves outcomes despite the often-advanced stage at presentation.
MET exon 14 skipping mutations have a distinctive patient profile - occurring in older patients, often with no smoking history - and correspondingly unique CT features: large tumors at presentation, frequent central necrosis, and a peripheral or pleural-based location. Bone metastases are common and frequently lytic (destructive rather than blastic), and brain metastases occur at rates comparable to EGFR mutations. The large, necrotic appearance of MET exon 14 tumors may reflect their tendency to grow rapidly while outpacing their blood supply.
BRAF mutations, which span multiple functional classes, lack a strongly distinctive radiological signature on CT, with imaging features varying substantially between BRAF V600E (the most actionable class) and other BRAF mutations. Studies show BRAF-mutant NSCLC tends toward solid peripheral tumors, but the radiological profile overlaps substantially with other driver subtypes, making CT imaging alone an unreliable predictor for this molecular subtype. Metastatic patterns are similarly variable.
HER2 mutations present with small, peripheral, spiculated tumors often with pleural tags, and miliary lung metastasis can occur similarly to EGFR-mutant disease, reflecting HER2's biological relationship to the same growth factor receptor family. NTRK fusions are so rare that no reliable CT imaging signature has been established, though the limited published cases suggest advanced lymph node involvement. For NTRK, molecular testing rather than imaging-based suspicion is the practical detection strategy.
Beyond qualitative radiologist-interpreted features, radiomic analysis extracts hundreds of quantitative texture and shape descriptors that can be systematically compared across molecular subtypes. For EGFR prediction, the most consistently identified radiomic features are related to tumor heterogeneity measures and density histogram features - solid versus ground-glass composition captured mathematically rather than visually.
Machine learning models combining radiomic features with clinical data achieve AUC values of 0.70 to 0.85 for EGFR prediction across multiple validation studies - better than chance but not definitive enough for clinical use without tissue confirmation. For ALK prediction, radiomic models show slightly lower performance (AUC 0.65 to 0.78), likely reflecting the more variable CT appearance of ALK-rearranged tumors. KRAS radiomic models show the weakest performance (AUC around 0.65), consistent with the less distinctive radiological phenotype.
Combining multiple data modalities improves performance for all drivers. Models that integrate CT radiomics with clinical features (age, sex, smoking status, tumor location) consistently outperform imaging-only or clinical-only models. For EGFR specifically, adding information from PET/CT metabolic features to CT radiomics yields further performance gains, and emerging work incorporating liquid biopsy circulating tumor DNA data with imaging creates multi-modal models approaching AUC values above 0.85.
The metastatic pattern data has immediate practical implications for staging workup design. Patients with CT features suggesting EGFR mutation should receive brain MRI as part of their initial staging given the high rate of brain metastasis, while patients with ALK-suggestive features warrant particular attention to the pericardium and unusual metastatic sites during staging CT review. Understanding driver-associated metastatic patterns can improve staging completeness before a molecular result is available.
During treatment with targeted therapies, knowledge of driver-specific metastatic tropism guides surveillance imaging choices. EGFR-mutant patients should have more frequent brain imaging during treatment and after progression, while RET-rearranged patients warrant more intensive bone surveillance. The emerging pattern of leptomeningeal progression in EGFR exon 20 patients could prompt earlier MRI spine inclusion in routine surveillance.
The radiogenomics findings also have implications for biopsy site selection when tissue is needed. If a patient presents with multiple potential biopsy targets - the primary tumor, a lymph node, and a bone lesion - understanding that different sites may have different mutation prevalence or may have undergone clonal evolution helps clinicians select the most informative target. In MET exon 14 disease, for example, biopsying the necrotic center of the tumor may yield non-representative acellular material.
Critical limitations constrain current radiogenomics evidence for all nine drivers. Most studies are retrospective, single-center, and use inconsistent CT acquisition protocols, making direct comparisons across studies difficult. Sample sizes for rare drivers such as RET, MET exon 14, BRAF, HER2, and NTRK are small - sometimes fewer than 50 cases - limiting statistical power and generalizability. Publication bias likely inflates reported performance metrics, as negative or weak radiogenomics findings are less likely to be published.
The field requires prospective multicenter validation studies that standardize CT acquisition, use consistent radiomic feature extraction pipelines, and prespecify the prediction model before analyzing results. Without such studies, the true clinical performance of CT-based driver prediction remains uncertain. Additionally, most radiogenomics studies focus on driver presence versus absence, without capturing the subtlety of different mutation variants (exon 19 deletion versus L858R in EGFR, for example) that may have different imaging correlates.
Despite these limitations, the radiogenomics approach shows sufficient biological rationale and early empirical support to justify continued development. The most plausible near-term clinical translation is not replacing molecular testing, but rather using imaging features to prioritize which patients should undergo expedited molecular testing, guide biopsy planning, and flag patients for additional staging procedures based on driver-specific metastatic risk profiles. Combining radiomic models with clinical risk factors and emerging liquid biopsy technologies represents the most promising path toward clinically implementable non-invasive driver characterization.