The Survival Gap by Stage. Lung cancer is the leading cause of cancer-related death in China and worldwide. The 5-year survival rate for stage IA1 patients reaches 95%, while stage IV patients have only 10% 5-year survival. This dramatic survival gap reflects a fundamental diagnostic challenge: the majority of patients present with advanced disease when surgical cure is no longer possible.
LDCT Screening Proven Effective. Major international randomized trials including the US National Lung Screening Trial (NLST) and the Dutch-Belgian NELSON trial have established that low-dose CT (LDCT) screening significantly reduces lung cancer mortality. A large-scale Chinese study of over 1 million participants further showed that even single-occasion LDCT screening reduces lung cancer deaths, providing a cost-effective strategy for resource-limited settings.
Limitations of Manual Image Interpretation. Traditional radiological assessment relies on clinician visual evaluation of qualitative features such as size, shape, enhancement patterns, and margin characteristics. This approach is time-consuming, subject to interobserver variability, and prone to missed or false diagnoses -- creating a clinical need for objective, quantitative, automated analysis tools.
Radiomics as a Solution. Radiomics, coined by Lambin and colleagues in 2012, systematically extracts large numbers of high-dimensional quantitative features from medical images including CT, MRI, and PET/CT. These features capture microscopic tumor heterogeneity invisible to the human eye and, combined with machine learning, provide new possibilities for early diagnosis and precision treatment of lung cancer beyond what traditional visual assessment can achieve.
Five-Step Radiomics Pipeline. The standard radiomics workflow consists of five sequential stages: image acquisition, region-of-interest (ROI) delineation, lesion segmentation, feature extraction and quantitative analysis, and model construction and validation for clinical prediction. Each stage introduces variability that must be controlled to ensure reproducible, generalizable models.
Traditional Challenges. Conventional radiomics faces several persistent limitations: manual ROI delineation by radiologists is time-consuming and experience-dependent, leading to segmentation inconsistencies; feature selection methods vary across studies; model generalizability is limited by acquisition parameter differences across scanners and institutions; and data standardization remains an unsolved challenge for multicenter applications.
Why AI Transforms Radiomics. The integration of artificial intelligence -- particularly deep learning -- addresses traditional radiomics limitations by enabling automated feature learning directly from image data without manual engineering. Deep learning models can analyze tumor volume dynamics, track multiple lesions simultaneously, interpret genotypic information from imaging phenotypes, and generalize across multiple disease contexts and imaging modalities.
AI Capabilities in Lung Cancer Imaging. AI-powered systems can generate high-quality CT images at reduced radiation doses, provide diagnostic decision support particularly for less experienced clinicians in resource-limited areas, and predict disease progression by comparing individual cases against large reference databases. An AI-assisted mobile CT screening program covering over 12,000 subjects achieved 95% nodule detection rate and 87% risk classification accuracy, with 85% of confirmed lung cancers identified at stage I.
The Nodule Management Challenge. LDCT screening reveals that 20 to 30% of participants have at least one lung nodule, generating substantial clinical uncertainty. Inaccurate nodule assessment risks unnecessary invasive procedures and patient anxiety on one hand, or missed cancers on the other. Radiomics provides objective malignancy probability estimates that improve on traditional qualitative assessment.
Radiomics Nodule Classification Performance. Multiple studies have demonstrated high accuracy for radiomics-based nodule classification: a nomogram built from 875 patients achieved AUC greater than 0.8 in training and validation; a specialized model for sub-centimeter solid nodules (under 1 cm) reached AUC 0.903; and a large-scale model validated across 16,797 nodules from four international screening studies achieved AUC 0.93, demonstrating population-level generalizability.
Differentiating Cryptogenic Mimics from Cancer. Pulmonary cryptococcosis granulomas can be radiologically indistinguishable from early lung cancer, leading to unnecessary surgery if misdiagnosed. A multicenter radiomics model distinguishing cryptococcal granulomas from lung adenocarcinoma achieved AUC 0.801, substantially exceeding junior radiologist performance (AUC 0.689). A deep learning local-global model combining nodule and whole-lung information reached AUC 0.88, outperforming local-only deep learning (AUC 0.84) and radiomics alone (AUC 0.79).
Histological Subtype Classification. Even among stage I patients, survival rates differ significantly between adenocarcinoma and squamous cell carcinoma, making noninvasive subtype classification clinically valuable. Deep learning-enhanced radiomics models for histological subtype classification of stage IA lung adenocarcinoma achieved validation AUC values ranging from 0.72 to 0.92 across multicenter testing, with overall accuracy of 83.5% versus 74.7% for conventional PET/CT and improved negative predictive value (95.6% vs. 88.2%).
Why Invasiveness Classification Matters. Accurate preoperative determination of adenocarcinoma invasiveness guides surgical decision-making: minimally invasive adenocarcinoma and adenocarcinoma in situ achieve near-100% 5-year survival and require only limited resection, while invasive adenocarcinoma warrants lobectomy with lymph node dissection. Approximately 50% of pure ground-glass nodules (pGGN) carry invasive potential, yet their CT features overlap substantially with benign lesions.
Radiomics Models for Invasiveness Prediction. Radiomics significantly improves invasiveness prediction beyond visual assessment. A CT feature-based risk model for invasive lung cancer demonstrated AUC of 0.910 with confirmed clinical utility by decision curve analysis. An integrated clinical-plus-imaging model achieved AUC greater than 0.85, substantially outperforming traditional models (AUC less than 0.73). A model simulating radiologist multi-reader adjudication demonstrated stable performance in multicenter environments.
Spread Through Air Spaces (STAS) Detection. The 2021 WHO lung cancer classification added spread through air spaces (STAS) as a pathological invasiveness criterion, but traditional STAS assessment is constrained by pathologist experience and tissue availability before resection. Quantitative imaging analysis using deep learning detection models can raise STAS identification accuracy to 93%, providing preoperative information previously inaccessible without tissue examination.
Future Model Integration. Current limitations in invasiveness prediction models include reliance on high-dimensional imaging features without incorporating clinician-style multidimensional reasoning. Future models should integrate radiomics quantitative features with traditional semantic imaging characteristics (nodule density, lobulation, pleural indentation) and clinical factors to more accurately simulate comprehensive clinical judgment and improve preoperative invasiveness assessment.
Recurrence Risk Despite Curative Surgery. Even after radical resection, 20 to 40% of stage I lung cancer patients experience recurrence, highlighting that pathological staging alone cannot fully capture biological risk. Imaging-based recurrence prediction before treatment completion could guide adjuvant therapy decisions and follow-up intensity.
Peritumoral Radiomics Improves Recurrence Prediction. A critical finding across recurrence prediction studies is that peritumoral (surrounding tissue) radiomics features outperform intratumoral features alone. Including expanded peritumoral regions improved recurrence prediction AUC to 0.73. Stability analysis revealed peritumoral features are more robust to scanning parameter variations than intratumoral features, providing a more reliable foundation for multicenter model generalization.
Adjuvant Chemotherapy Selection. Current guidelines do not recommend adjuvant chemotherapy for stage IA NSCLC after curative resection, and its benefit in IB disease remains debated. CT-based radiomics risk scores can identify high-risk stage II patients who will benefit most from chemotherapy, while simultaneously identifying low-risk patients who can be spared toxic treatment. They can also identify stage IA patients with sufficiently high biological risk to warrant adjuvant therapy consideration despite guideline recommendations against it.
Immune Checkpoint Inhibitor Response Prediction. Ensemble deep learning CT-based models have demonstrated ability to predict benefit from immune checkpoint inhibitors in NSCLC patients in retrospective studies (published in Lancet Digital Health), enabling prospective selection of immunotherapy responders before treatment begins. This extends radiomics application from anatomical characterization to functional immune landscape inference.
Three Key Barriers to Clinical Translation. Despite substantial promise, three fundamental challenges prevent widespread clinical deployment of radiomics in lung cancer: non-standardized image acquisition parameters across institutions make results difficult to compare and integrate; data privacy and ethics concerns in large-scale medical data applications require regulatory frameworks; and the black-box nature of deep learning limits clinician understanding and trust in model decision-making.
Standardization Efforts Underway. The European Society of Radiology published a 58-item evaluation standard for radiomics research in 2023 (CLEAR checklist), providing a normalization framework for future studies. Harmonization techniques such as Combat reduce between-scanner variability and improve multicenter model performance (AUC 0.9 for malignancy classification). Thin-slice CT substantially improves early cancer detection compared to thick-slice CT (83.1% vs. 52.1% for stage IA1 detection).
Explainability as a Clinical Priority. The inability of deep learning models to explain their predictions in human-interpretable terms is a major adoption barrier for clinical radiologists and oncologists who must justify diagnostic and treatment decisions. Future research must prioritize model interpretability through attention visualization, saliency mapping, and integration with established radiological knowledge to build the clinician trust needed for routine clinical deployment.
Research Priorities for the Next Decade. Future advances require multicenter prospective validation studies, enhanced model interpretability, and integration into clinical workflows rather than as standalone tools. Combining radiomics with multi-omics data (genomics, proteomics) represents a promising direction for more comprehensive precision medicine applications. The field's maturation from single-center research to clinical-grade tools requires multidisciplinary collaboration among radiologists, oncologists, data scientists, and regulatory experts.