Clinical need for better AI tools. CT is central to lung cancer screening, diagnosis, staging, and treatment evaluation. Deep learning models have shown promise for quantitative analysis of CT scans -- predicting lymph node metastasis and prognosis -- but widespread clinical adoption remains limited by several fundamental barriers: low reproducibility due to scanner variability, high cost of expert-annotated datasets, inconsistent imaging quality, and the single-task focus of most models that cannot address the multidimensional complexity of lung cancer care.
Limitations of supervised learning in small-sample settings. Traditional supervised learning for CT-based lung cancer tasks requires labor-intensive region-of-interest annotations by expert radiologists. Small labeled datasets increase overfitting risk, and models trained at single institutions often fail to generalize across scanner vendors, acquisition protocols, and patient populations. Most existing approaches produce single-purpose models rather than unified frameworks that can address multiple clinical questions simultaneously.
Self-supervised learning as the solution. Self-supervised learning (SSL) enables models to learn meaningful feature representations from large volumes of unlabeled data, then adapt those representations to specific downstream tasks with minimal labeled examples. SSL-based foundation models have demonstrated performance matching or exceeding fully supervised models using as little as 1% of annotated data in cancer prediction tasks. However, prior SSL foundation models for CT imaging have struggled with interpretable three-dimensional analysis of thoracic oncology due to architectural and data management challenges.
UCLIF: a unified CT foundation model. This study developed the Unified CT-Based Lung Cancer Imaging Foundation (UCLIF) model -- an SSL framework pretrained on 33,901 three-dimensional chest CT scans and then fine-tuned across four clinically critical downstream tasks: histologic subtype classification, cancer staging, survival prediction, and recurrence prediction. The model attends to lesion regions, whole-lung structure, and peritumoral context to extract comprehensive representations relevant across the full cancer care continuum.
Stage I: SSL pretraining on 33,901 CT scans. UCLIF was pretrained using 33,901 chest CT scans from 27,092 adult patients drawn from 15 publicly available Cancer Imaging Archive (TCIA) datasets. These datasets span diverse imaging modalities including low-dose CT, spiral CT, four-dimensional fan-beam CT, and four-dimensional cone-beam CT, covering multimodal approaches integrating imaging with genomics, proteomics, and pathology. No patient labels were used during pretraining; the model learned representations entirely through a contrastive masked image modeling task.
UCLIF architecture: Vision Transformer autoencoder. The model adopts a Vision Transformer encoder paired with a complementary Transformer decoder. During pretraining, the encoder receives image patches with selective masking and generates advanced feature representations. The decoder reconstructs obscured portions of the scan from masked tokens using encoder outputs as context -- a self-supervised image completion strategy that forces the model to learn robust visual representations without explicit labels. This architecture is specifically designed for the three-dimensional challenges of volumetric CT analysis.
Data preprocessing and peritumoral feature extraction. CT volumes were resampled to isotropic 1 mm x 1 mm x 1 mm voxel size using cubic B-spline interpolation, then intensity-normalized to range [0,1] within a -1400 to 400 HU window. Two thoracic radiologists (each with 10 years' experience) delineated lesion boundaries by consensus, with senior adjudication for discrepancies. Surrounding tissue was evaluated within an 8 mm peritumoral ring, enabling the model to capture both intratumoral and peritumoral contextual features relevant to staging and prognosis.
Stage II: fine-tuning for four downstream clinical tasks. The pretrained encoder was extended with a classification module using PCA compression (from 2048 to 256 dimensions) and fully connected layers for binary classification tasks. Four task categories were evaluated: histologic subtype classification (adenocarcinoma, large cell lung cancer, squamous cell carcinoma), TNM staging (stages I to IV), survival prediction (1-, 3-, and 5-year outcomes), and recurrence prediction. Datasets for fine-tuning were strictly separated from pretraining data to prevent data leakage, with training on NSCLC Radiogenomics data and testing on independent TCGA-LUSC and other datasets.
Histologic subtype classification results. UCLIF achieved AUC 0.96 (95% CI: 0.88 to 1.00) for adenocarcinoma, AUC 0.82 (95% CI: 0.60 to 0.98) for large cell lung cancer, and AUC 0.93 (95% CI: 0.80 to 0.99) for squamous cell carcinoma. Overall accuracy across the three subtypes was 63% for adenocarcinoma, 26% for large cell lung cancer, and 79% for squamous cell carcinoma. Squamous cell carcinoma classification showed particularly strong specificity (92%) and positive predictive value (75%). All pairwise comparisons with SSL-Lung and SSL-ImageNet baselines were statistically significant (DeLong test, p less than 0.001).
The large cell lung cancer challenge. The lower performance for large cell lung cancer (AUC 0.82 and accuracy 26%) reflects both limited sample size in this class and class imbalance in the available labeled datasets. This result is consistent with prior literature on rare lung cancer subtypes and does not reflect a general weakness of the architecture. All other subtypes showed consistently strong metrics, indicating the UCLIF model can manage complex multiclass tasks under realistic imbalanced data conditions.
TNM staging classification. For cancer staging across the four TNM stages, UCLIF achieved AUC 0.95 (95% CI: 0.79 to 1.00) for stage I, AUC 0.99 (95% CI: 0.96 to 1.00) for stage II, AUC 0.92 (95% CI: 0.74 to 1.00) for stage III, and AUC 0.91 (95% CI: 0.78 to 1.00) for stage IV. Accuracy per stage was 89%, 86%, 72%, and 95% for stages I to IV. These results significantly outperformed both SSL-Lung and SSL-ImageNet baselines (p less than 0.001) and exceeded a prior ResNet-50-based model restricted to T category classification -- demonstrating the advantage of a unified framework covering all stages simultaneously.
Why domain-specific pretraining outperforms generic SSL. The consistent superiority of UCLIF over SSL-ImageNet -- which was pretrained on natural photographs rather than CT scans -- demonstrates that generic visual representations fail to capture tumor-specific structural and textural features in three-dimensional thoracic imaging. Domain-specific SSL pretraining on chest CT yielded more informative and transferable representations for thoracic oncology, consistent with a growing body of medical imaging literature showing medical-specific pretraining substantially improves disease classification.
Survival prediction across three time horizons. UCLIF achieved AUC 0.97 (95% CI: 0.92 to 1.00) for 1-year survival prediction, AUC 0.90 (95% CI: 0.72 to 0.98) for 3-year survival, and AUC 0.90 (95% CI: 0.77 to 1.00) for 5-year survival, along with accuracy of 91%, 88%, and 91% respectively. These results were tested on the independent TCGA-LUSC dataset, which differed from the training cohort in acquisition protocols, scanner vendors, and patient population, making this a genuine external validation of prognostic performance.
Recurrence prediction with high specificity. For recurrence prediction, UCLIF achieved accuracy 94% (81/86), AUC 0.95 (95% CI: 0.88 to 0.99), specificity 99%, negative predictive value 95%, and positive predictive value 83%. This means the model almost never falsely predicts recurrence in patients who will not recur -- a critical clinical property because false positive recurrence predictions would trigger unnecessary interventions. Performance exceeded prior single-hospital prognostic models (which reported AUC 0.860) and early-recurrence models using ResNet18 (AUC 0.817).
Multivariable Cox and DeepSurv model comparison. Multivariable Cox proportional hazards models incorporating UCLIF risk scores alongside baseline covariates (age, sex, smoking status) confirmed that high-risk patients faced significantly greater mortality and recurrence risk than low-risk patients in long-term analyses. The 1-year multivariate result was not statistically significant, likely reflecting peritreatment factors not captured by baseline CT alone. DeepSurv models produced comparable but slightly inferior discrimination and calibration compared to the Cox-based UCLIF pipeline under the same data splits.
Outperforming all comparison algorithms. UCLIF consistently exceeded performance of convolutional neural networks (DenseNet121, ResNet34, ShuffleNet), conventional machine learning classifiers (logistic regression, SVM, random forest, XGBoost), SSL-Lung (domain-specific SSL on lung scans only), and SSL-ImageNet (generic natural image SSL) across all tasks and datasets. The UCLIF model achieved higher accuracy and AUC than comparison methods at p less than 0.001 across multiple datasets, with consistently narrower confidence intervals indicating greater model stability.
Reducing annotation burden. The UCLIF model's most significant practical advantage is its ability to achieve competitive performance using substantially smaller labeled datasets than supervised learning requires. By leveraging 33,901 unlabeled CT scans for pretraining, the model learns rich visual representations that transfer efficiently to downstream tasks with limited labeled examples. This is particularly important in lung cancer imaging where acquiring expert-annotated three-dimensional CT volumes is costly and time-consuming.
Robustness to real-world imaging heterogeneity. UCLIF demonstrated substantial robustness to domain shifts across external validation datasets that differed in acquisition protocols, scanner vendors, reconstruction parameters, and patient populations. Conventional supervised models trained from scratch or initialized on natural images showed performance deterioration and variability across external test sets, while UCLIF maintained consistently superior and stable performance. This robustness is a direct consequence of pretraining on diverse multi-institutional CT data that forced the model to learn acquisition-invariant representations.
A unified framework for multiple clinical tasks. Unlike prior approaches that produce single-purpose models, UCLIF addresses histologic classification, staging, survival, and recurrence within a single unified framework. This multi-task architecture supports more comprehensive clinical decision support -- a clinician could use a single model from initial CT-based subtyping through postoperative recurrence surveillance, rather than deploying separate specialized tools for each question. This capability supports personalized, data-driven management across the full lung cancer care continuum.
Comparison with existing foundation models in medicine. UCLIF extends the concept of SSL foundation models from retinal imaging and chest radiographs into three-dimensional thoracic oncology CT -- a more complex domain requiring volumetric analysis, peritumoral context, and simultaneous discrimination of subtle differences in staging and prognosis. Its superior performance over SSL-Lung (which also used lung CT but without the contrastive masked image modeling pretraining task) demonstrates that the specific SSL methodology matters, not just the data domain.
CT-only data input. The current UCLIF model relies solely on CT images without incorporating complementary clinical, pathologic, or molecular data. Integration of pathology, genomics, and clinical variables alongside imaging could substantially improve predictive performance and provide more comprehensive decision support. This represents the most important next step in model development.
Limited interpretability. Like most deep learning models, UCLIF functions as a black box with limited interpretability, which may hinder clinician trust and adoption. The authors identify attention-based visualizations, integrated gradients, and perturbation or occlusion tests as specific explainability methods that should be incorporated in future iterations to enable clinicians to understand which imaging features drive specific predictions.
Retrospective design and lack of prospective validation. All datasets were retrospective, and the model has not been prospectively validated or embedded into routine clinical workflows. Performance in prospective deployment settings -- where patients are consecutive, follow-up is standardized, and CT protocols are controlled -- may differ from the retrospective multicenter results reported here. Residual biases from site-specific referral patterns and imaging protocols cannot be excluded.
Priority future directions. The authors call for large-scale prospective multicenter studies with consecutive patient enrollment, prespecified operating thresholds, and workflow-embedded deployment to validate clinical utility. Integration with triage pipelines, presurgical planning systems, and multidisciplinary team discussion platforms represents the path from research proof-of-concept to routine clinical impact. Multimodal data fusion incorporating genomic, pathologic, and clinical variables alongside CT remains the most promising technical direction for performance improvement.