Surgery is the primary curative treatment for early-stage non-small cell lung cancer (NSCLC), which makes up about 85% of all lung cancers. When NSCLC is caught early, before it has spread to lymph nodes or distant organs, surgical removal of the tumor offers the best chance of cure. Advances in low-dose CT screening have made earlier detection increasingly possible.
Despite successful surgery, 30-55% of patients experience cancer recurrence within five years. This alarmingly high rate persists even when the surgeon achieves clear margins and the pathology report shows no residual cancer. Two patients who appear identical by conventional staging may have dramatically different outcomes - one may be cured while the other recurs within two years.
Adjuvant therapies like chemotherapy and radiotherapy after surgery do not reliably prevent recurrence. The toxicity of chemotherapy itself contributes to post-surgery mortality, and survival advantages from adjuvant treatment are not consistently observed across patients. There is genuine controversy about who should receive adjuvant treatment and what form it should take.
The core problem is that current staging systems are insufficient to predict individual outcomes. The TNM staging system (Tumor, Node, Metastasis) describes anatomical extent of disease but cannot capture molecular heterogeneity - two stage I tumors can have very different genetic profiles, microenvironments, and aggressiveness. Better tools for predicting recurrence risk at the individual patient level are urgently needed.
Radiomics, machine learning, and deep learning are the three main AI approaches being applied to recurrence prediction. Radiomics extracts hundreds or thousands of quantitative measurements from medical images - describing tumor texture, shape, intensity distribution, and spatial relationships - and links them statistically to outcomes. Machine learning uses algorithms that identify patterns in these features and other clinical data. Deep learning uses neural networks to learn features directly from the raw image data without manual extraction.
CT (computed tomography) is the most commonly used imaging source for radiomics in this setting. It provides detailed three-dimensional anatomical information about tumor size, shape, density, and relationship to surrounding structures. PET (positron emission tomography) adds metabolic information - it shows how actively tumor cells are consuming glucose, which relates to cell proliferation rate and aggressiveness. The combination of both (PET/CT) provides complementary information that neither modality alone can deliver.
The key insight driving this research is that tumors contain hidden biological information visible in their imaging appearance. The specific texture of a tumor on CT - whether it is homogeneous or heterogeneous, solid or part-solid, smooth-edged or spiculated - reflects underlying biology such as genetic mutations, cell density, and microenvironmental composition. Radiomics attempts to extract and quantify these visual patterns in a reproducible, objective way.
AI models can potentially improve on TNM staging by incorporating this imaging information. Multiple studies reviewed here showed that AI models outperformed stage-only prediction, sometimes substantially. The goal is not to replace clinician judgment but to provide a more precise, data-driven risk estimate that guides treatment decisions - particularly around which patients need aggressive adjuvant treatment versus watchful waiting.
The researchers conducted a structured literature review of studies published between 2018 and 2024. The search covered Google Scholar, PubMed, and Scopus using specific terms combining NSCLC, recurrence prediction, AI methods, and imaging modalities. Starting from 162 identified articles, the team applied systematic eligibility criteria to arrive at 16 studies meeting all requirements.
Strict inclusion criteria focused on the most clinically relevant scenario. Only studies of early-stage NSCLC treated with surgery, using AI methods applied to preoperative CT or PET imaging (with or without clinical data), and predicting recurrence risk (not recurrence location or metastasis) were included. Studies using genomic data or pathological tissue images were excluded, keeping the focus on what can be gathered non-invasively before surgery.
Study quality was assessed using the METRICS tool (METhodological RadiomICs Score). This is a validated quality scoring instrument endorsed by the European Society of Medical Imaging Informatics. It evaluates 30 items across nine categories including study design, imaging data quality, segmentation methods, feature processing, and open science practices. This allowed the reviewers to rate each study as excellent, good, or moderate quality.
The 16 included studies varied considerably in design, modalities, and AI methods. Most were retrospective single-center studies. Four used CT alone, eight used CT with clinical data, one used PET alone, and three combined PET, CT, and clinical data. AI approaches ranged from traditional statistical analysis of radiomics features to state-of-the-art deep learning architectures including convolutional neural networks and vision transformers.
Studies combining CT radiomics with clinical data consistently outperformed either approach alone. Wang et al. (2024) achieved an impressive AUC (area under the ROC curve) of 0.937 for their combined radiomics-clinical model, compared to only 0.788 for the clinical-only model. D'Antonoli et al. (2020) showed similar improvement: the combined model achieved AUC 0.760 versus 0.24 for clinical factors alone - a dramatic difference that underscores how much additional prognostic information is embedded in CT images beyond what clinical variables capture.
Including the peritumoral region (the lung tissue immediately surrounding the tumor) improved predictions. D'Antonoli et al. specifically studied both the tumor itself and the surrounding lung parenchyma, finding that combining both regions outperformed using the tumor alone. This makes biological sense - tumors modify their microenvironment, and features of the surrounding tissue may reflect invasive potential or immune response.
CT image normalization - standardizing pixel intensity values across scans from different scanners and protocols - significantly improved model performance. Park et al. (2022) showed that more radiomic features survived quality filtering after normalization, and 3-year recurrence-free survival prediction improved measurably when normalized CT images were used. This highlights a practical challenge: raw LDCT images from different clinical sites are not directly comparable without preprocessing.
Random forest classifiers were among the most commonly used and best-performing machine learning approaches. Wang et al. (2019) achieved 84.7% accuracy and AUC of 0.88 using random forest applied to CT grayscale, shape, and texture features. This well-established machine learning algorithm proved effective for combining multiple radiomics features into a recurrence risk score.
PET-based radiomics outperformed standard PET metrics for predicting recurrence. Ahn et al. (2019) showed that texture features from PET (describing patterns of how metabolic activity is distributed within the tumor) outperformed conventional PET measures like standardized uptake value (SUV), metabolic tumor volume, and total lesion glycolysis. Their random forest model achieved an AUC of 0.956 - one of the highest values in the entire review - using only PET texture features.
Adding PET data to CT radiomics and clinical information generally improved prognostic accuracy. Christie et al. (2022) found that their multimodality model (PET + CT + clinical data including stage) achieved a concordance index of 0.76 compared to 0.60 for stage alone. While the absolute improvement varied across studies, the trend was consistent: more complementary information sources generally produced better predictions, especially when models were validated in external cohorts.
Interestingly, simply adding clinical variables to imaging-only radiomics models did not always help. Kirienko et al. (2018) found that adding age, sex, histology, and stage to PET-only or CT-only radiomics models did not improve their already-predictive radiomic signatures. This suggests that the radiomics features may already capture information that overlaps with standard clinical variables, and that the genuine added value of combining modalities comes when the information is truly complementary.
Christie et al. (2021) specifically studied multimodal tumor/peritumoral features from both PET and CT. Their combined approach achieved a C-index of 0.81 in training and 0.79 in testing - strong performance for a prognostic model. Importantly, the multimodal model was significantly better at stratifying patients into low and high-risk groups than either single modality or stage alone.
Deep learning models - particularly convolutional neural networks (CNNs) - showed strong performance in predicting recurrence from CT images. Fanizzi et al. (2023) compared multiple deep learning architectures including InceptionV3, ResNet50, DenseNet, and several types of Vision Transformers. The best performer was InceptionV3 CNN, achieving AUC 0.91, accuracy 0.89, sensitivity 0.85, and specificity 0.90 - better than any radiomics approach in the same study.
Transfer learning helped overcome the challenge of small medical imaging datasets. Several studies used CNNs pre-trained on large natural image datasets (like ImageNet) and then fine-tuned them on their NSCLC data. This transfer of knowledge from general visual tasks to the specific problem of lung cancer recurrence prediction allowed models to achieve reasonable performance even with fewer than 200 training patients - far fewer than would be needed to train a deep learning model from scratch.
A sophisticated Transformer-Graph model by Lian et al. (2022) predicted 5-year outcomes in the largest dataset reviewed (1,705 patients). This approach combined a Vision Transformer (which processes CT image features) with a Graph Neural Network (which encoded relationships between patients sharing similar clinical characteristics). The combined model achieved an RFS AUC of 0.726 in testing and 0.700 in external validation, outperforming both TNM staging alone and a standard ResNet imaging model.
CNNs generally outperformed Vision Transformers for NSCLC recurrence prediction. While Vision Transformers have achieved state-of-the-art results in natural image tasks and some medical imaging applications, CNNs maintained an edge in this specific context - likely because the relatively small dataset sizes favor the better-established spatial feature extraction capabilities of CNNs over the data-hungry transformer attention mechanisms.
Small sample sizes are the most pervasive limitation in this field. Most studies included fewer than 200 patients, with only one (Lian et al.) exceeding 1,000. Small datasets increase the risk of overfitting - where a model learns to memorize the training data rather than discover generalizable patterns - and make it difficult to train complex deep learning architectures reliably. High-performing metrics on small datasets should be interpreted with caution.
Only 3 of the 16 reviewed studies included external validation. External validation - testing a model on a completely independent dataset from a different institution - is the gold standard for assessing clinical generalizability. Without it, a model's apparent performance may reflect characteristics specific to one hospital's scanner, patient population, or imaging protocol rather than true predictive power. The rarity of external validation in this literature is a major gap.
Almost all studies were retrospective, introducing selection bias. Retrospective studies analyze data that was collected for other purposes, which can introduce systematic biases - for example, if high-risk patients were selected for surgery in ways that correlate with radiomics features that predict recurrence. Prospective studies, where patients are enrolled specifically for the research purpose and imaged according to a standardized protocol, are needed to confirm findings.
Lack of standardization in preprocessing, segmentation, feature extraction, and evaluation metrics makes comparing across studies very difficult. One study might manually segment tumors while another uses semi-automatic methods; one might use PyRadiomics for feature extraction while another uses proprietary software. These methodological differences mean that a model achieving AUC 0.90 in one study is not directly comparable to a model achieving AUC 0.80 in another.
Multimodal approaches combining imaging and clinical data consistently outperform single-modality models and represent the most promising direction. Future research should prioritize developing and validating multimodal frameworks that integrate CT texture features, PET metabolic features, and clinical variables including TNM stage, histology, and patient demographics into robust, well-calibrated risk scores.
Larger, prospective, multicenter studies are essential for establishing clinical utility. Such studies would need standardized imaging protocols, consistent tumor segmentation, and pre-registered analysis plans to ensure results are not optimized to look good retrospectively. Ideally, they would also include health economic analyses to determine whether AI-guided treatment stratification actually improves outcomes cost-effectively.
Interpretability must improve before widespread clinical adoption. Many deep learning models remain 'black boxes' - they produce predictions without explanations clinicians can evaluate. Techniques like explainable AI, which highlight which image features drove a particular prediction, and radiomics-guided deep learning (which anchors neural network features to established radiomic interpretations) are promising directions for making AI predictions more trustworthy and actionable.
The clinical payoff from solving these challenges would be substantial. An AI model that reliably identifies the 30-55% of early-stage NSCLC patients who will relapse could transform post-surgical management: high-risk patients could receive aggressive adjuvant therapy and closer surveillance, while low-risk patients could be spared toxic treatments. This kind of precision oncology decision-making is exactly what AI in medical imaging is ultimately positioned to enable.