What are ground-glass nodules? Ground-glass nodules (GGNs) are hazy areas of increased density visible on CT scans that do not obscure the underlying blood vessels or airways. They are increasingly detected thanks to widespread low-dose CT screening and represent one of the earliest imaging signs of lung adenocarcinoma.
Two main types exist. Pure ground-glass nodules (pGGNs) show uniform haziness without any solid component and most often correspond to pre-invasive or minimally invasive tumors. Mixed GGNs contain both hazy and solid areas and are more frequently associated with invasive adenocarcinoma, making this distinction clinically important for treatment planning.
Traditional reading has real limits. Radiologists historically assessed GGNs by measuring size, shape, spiculation, and the ratio of solid to ground-glass components. While useful, these methods suffer from subjectivity and inconsistency between readers, especially for small or subtle nodules.
Why this matters for patients. GGNs provide a rare window onto the very earliest stages of lung cancer development. Accurately characterizing them - and connecting their appearance to underlying genetic changes - could allow treatment to begin before invasion occurs, significantly improving survival.
Radiomics goes far beyond what the eye can see. Radiomics is a technique that extracts hundreds to thousands of quantitative features from CT images - including texture, shape, and intensity patterns - that no human reader could systematically measure. These features are then fed into machine learning models to predict malignancy, invasiveness, or molecular status.
Deep learning automates the hardest steps. Convolutional neural networks (CNNs), especially 3D architectures like U-Net, can automatically detect and outline GGNs in CT volumes with high accuracy. This removes the tedious and variable step of manual segmentation and ensures consistent feature measurement across patients and institutions.
Feature selection is essential. Because radiomics can generate thousands of candidate features, methods like LASSO regression and minimum redundancy maximum relevance (mRMR) are used to identify the small subset that genuinely predicts clinical outcomes, reducing noise and preventing overfitting in predictive models.
Limitations of traditional approaches drove this shift. Manual imaging assessment is limited by inter-observer variability, inability to quantify subtle textural differences, and difficulty tracking nodule evolution over time. These shortcomings motivated the development of automated, reproducible AI-based pipelines that can capture tumor heterogeneity more completely.
EGFR is the most common driver mutation. Mutations in the epidermal growth factor receptor (EGFR) gene - particularly exon 19 deletions and the L858R substitution - account for up to 70% of lung adenocarcinoma cases in Asian populations. These mutations cause the receptor to be permanently switched on, driving uncontrolled cell growth and making the tumor sensitive to targeted drugs called tyrosine kinase inhibitors (TKIs).
Other important mutations include ALK, KRAS, and ROS1. ALK gene rearrangements occur in about 2.5 to 6.6% of patients and are more common in younger non-smokers. KRAS mutations (6-16% of cases) tend to occur in male smokers and are linked to more aggressive disease. ROS1 rearrangements are rarer still but similarly drive constitutive kinase signaling.
Each mutation shapes tumor biology differently. EGFR mutations are linked to well-differentiated, non-invasive tumors and better prognosis. KRAS mutations tend to produce poorly differentiated, aggressive cancers. ALK and ROS1 rearrangements create fusion proteins that continuously activate growth signals and respond well to specific inhibitor drugs.
Immune microenvironment differs by mutation type. KRAS-mutant GGNs show higher immune cell infiltration and stronger immune inhibitory signals compared to EGFR-mutant nodules, potentially explaining their different responses to immunotherapy. This biology shapes not just imaging appearance but also treatment strategy.
EGFR mutations leave visible CT fingerprints. Nodules harboring EGFR mutations tend to be larger and more frequently show the vacuole or honeycomb sign on high-resolution CT. Ground-glass opacity or mixed ground-glass appearance is also more common in EGFR-mutant tumors, particularly in non-smoking women with multiple lung adenocarcinomas.
Quantitative CT parameters correlate with mutation status. Three-dimensional measurements including bronchial sign, pleural indentation, vascular bundle sign, maximum diameter, nodule volume, average CT value, and the proportion of solid component all correlate significantly with EGFR mutations and ALK rearrangements, supporting the idea that imaging can serve as a non-invasive molecular readout.
Mutation patterns shift as nodules progress. Integrated genomic studies show that EGFR mutations are early events in GGN development, followed by additional mutations in RBM10 or TP53 as tumors progress from adenocarcinoma in situ to invasive cancer. Tumor mutation burden and genomic complexity increase step by step along this continuum.
88% of GGNs carry driver pathway mutations. A large study of 334 resected GGN nodules found that 88% harbored at least one mutation in the RTK/RAS signaling pathway. Importantly, different nodules within the same patient often carried different, mutually exclusive driver mutations, highlighting the complexity of managing multiple simultaneous GGNs.
Two main fusion approaches exist. Feature-level fusion combines raw features from different data sources (CT images, genomic data, clinical variables) into a single unified representation before training a model. Decision-level fusion instead trains separate models on each data type and then combines their outputs using voting or ensemble methods. Both approaches have demonstrated value in lung adenocarcinoma research.
Combining data types outperforms any single source. A multimodal model integrating CT images, patient demographics, and serum tumor markers using a residual learning architecture achieved 88.5% accuracy and an AUC of 0.957 for distinguishing invasive adenocarcinoma from non-invasive GGNs - outperforming senior radiologists working from images alone.
Graph neural networks add a new dimension. Graph neural networks (GNNs) represent clinical, imaging, and semantic data as interconnected nodes and edges, capturing relationships between data types that simpler fusion methods miss. This approach improved both performance and interpretability in lung adenocarcinoma classification tasks.
Predicting EGFR status without biopsy. Fusion of radiomic and deep learning features extracted from both the tumor core and surrounding tissue achieved AUCs up to 0.925 internally and 0.889 externally for predicting EGFR mutation status - performance levels that could meaningfully reduce the need for invasive tissue sampling in some patients.
AI dramatically improves detection sensitivity. Large-scale clinical trials embedding AI into CT screening workflows have reported sensitivity improvements of up to 20.7%, with overall detection rates rising from approximately 67.7% to 88.4% for various nodule sizes including those traditionally difficult for radiologists to spot. Some deep learning models have achieved near-99% accuracy in GGN detection.
AI goes beyond detection to risk stratification. AI-based radiomics can predict early recurrence risk in stage IA lung cancer using imaging features like solid part size and volume ratios. Quantum machine learning-integrated models have achieved up to 89.23% accuracy in classifying GGNs, enabling earlier identification of patients who need more aggressive management.
Non-invasive molecular profiling becomes possible. AI models trained on CT images have achieved AUC values up to 0.91 for predicting EGFR mutation status, potentially allowing oncologists to identify targeted therapy candidates without waiting for biopsy results. This is especially valuable when tissue is limited or biopsy carries high risk.
Access challenges can be addressed. Mobile low-dose CT units equipped with AI diagnostic systems have shown feasibility in rural and underserved settings, achieving lung cancer detection rates around 0.7% while reducing health disparities. AI also shortens detection times and improves concordance compared to manual reading, supporting wider screening coverage.
AI refines surgical planning decisions. Deep learning models can classify GGNs as atypical adenomatous hyperplasia, adenocarcinoma in situ, minimally invasive adenocarcinoma, or invasive adenocarcinoma based on CT features with high accuracy. This stratification directly informs whether a patient needs sublobar resection, full lobectomy, or watchful waiting - decisions that greatly affect quality of life.
Prognostic models combine imaging with molecular data. Nomograms integrating novel biomarkers from RNA sequencing with CT imaging features have demonstrated AUCs exceeding 0.8 for predicting outcomes in GGN-associated adenocarcinoma. Radiomics-based models reached AUCs up to 0.92 in distinguishing invasive from pre-invasive lesions in independent validation cohorts.
Targeted therapy selection benefits from AI. AI algorithms that infer EGFR mutation status from CT radiomics signatures can help identify candidates for tyrosine kinase inhibitors early in the diagnostic workup, potentially shortening the time from detection to treatment initiation. AI also shows promise for predicting immune microenvironment characteristics relevant to immunotherapy decisions.
Challenges remain in translating research to clinics. Many prognostic AI models are built on retrospective single-center data and lack external validation, limiting their generalizability. The opaque decision-making of deep learning systems also raises accountability concerns, requiring explainable AI frameworks and multidisciplinary oversight before widespread clinical adoption.
Heterogeneous data is a major obstacle. AI models trained on CT images are sensitive to differences in scanning protocols, tube voltage, reconstruction algorithms, and slice thickness. Studies have shown these technical variables directly affect measured nodule size and density - the very inputs that AI models depend on - making it difficult to deploy models trained at one institution across others.
Annotation consistency is critical but hard to achieve. Even expert radiologists disagree on the boundaries and classification of GGNs, especially for small or part-solid nodules. When training datasets contain inconsistent labels, AI models learn from noise and may perform poorly on new cases. Standardized annotation protocols and quality control are essential.
Data augmentation and transfer learning help with small datasets. Because well-annotated GGN datasets are rare, researchers apply rotation, scaling, and noise addition to artificially expand training sets. Transfer learning reuses models pretrained on large general datasets and fine-tunes them for GGN-specific tasks, reducing the number of labeled examples required.
Multi-institutional collaboration is the long-term solution. Establishing large, harmonized databases with standardized imaging protocols and annotation criteria across institutions is fundamental. Advanced reconstruction methods like spectral photon-counting CT offer improved nodule characterization that could make AI inputs more consistent across sites.
The black box problem limits clinical trust. Most deep learning models cannot explain why they reach a particular conclusion, making clinicians reluctant to act on AI recommendations for high-stakes decisions like surgery or biopsy. Regulatory bodies also require clear understanding of model behavior before approving AI diagnostic tools.
Explainable AI techniques are being developed. Saliency maps and heatmaps can highlight which regions of a CT image drove the AI's prediction, allowing radiologists to verify that the model is focusing on relevant features like nodule size, density, or spiculation rather than imaging artifacts. SHAP values quantify each feature's contribution to individual predictions.
Bias and equity are ethical priorities. AI models trained predominantly on data from specific ethnic groups may perform poorly for underrepresented populations. Since EGFR mutation rates and GGN imaging characteristics differ across ethnicities, models must be evaluated and validated across diverse populations to avoid amplifying existing health disparities.
Federated learning offers a privacy-preserving path forward. Federated learning trains AI models across multiple institutions without transferring raw patient data, allowing collaboration while preserving privacy and complying with data protection regulations. This approach is particularly important for sensitive genomic and imaging data used in lung cancer AI research.
Multi-omics integration is the next frontier. Future AI systems will need to integrate CT imaging, genomic sequencing, transcriptomics, proteomics, and liquid biopsy data into unified models. This comprehensive multi-omics approach promises to capture tumor biology more completely than any single data type, enabling more accurate early diagnosis and treatment personalization.
Prospective validation is urgently needed. The vast majority of current AI studies in GGN analysis are retrospective and single-center. Large-scale prospective clinical trials that follow patients over time are essential to confirm that AI predictions translate into genuine improvements in survival and quality of life rather than simply better test performance metrics.
The imaging-genomics bridge has transformative potential. By systematically linking CT imaging phenotypes with driver gene mutations through AI, clinicians could one day determine a tumor's molecular profile from a scan taken minutes after detection - potentially eliminating weeks of waiting for biopsy and genomic sequencing results while guiding therapy selection.
Human-AI collaboration remains the model. The review emphasizes that AI should augment rather than replace clinical judgment. Multidisciplinary teams incorporating oncologists, radiologists, molecular biologists, and AI scientists working with interpretable, validated AI tools represent the most realistic path to realizing the precision oncology vision for lung adenocarcinoma presenting as GGNs.