Oligometastatic disease is defined as 1 to 5 metastatic lesions in the body where aggressive local treatment can be applied to all sites. Rather than systemic chemotherapy alone, combining stereotactic body radiotherapy (SBRT) with systemic treatment has been shown to confer a survival benefit in this setting.
The lung is the most common site of distant metastasis for many solid tumors, making it a frequent target for SBRT. Published data show median 1-year and 5-year local control rates of approximately 90% and 79% after lung SBRT, respectively - demonstrating the treatment's effectiveness when applied appropriately.
Despite these good overall control rates, some lesions do recur locally, and patient- or tumor-specific factors alone have proven insufficient to fully predict which lesions will fail. Identifying those at higher risk of local recurrence ahead of time could allow for personalized treatment intensification.
Radiomics - the extraction of large numbers of quantitative features from medical images - offers a promising complement to clinical data for this purpose. However, its application in lung oligometastases treated with SBRT remains limited, motivating this study.
This retrospective study analyzed 80 lung oligometastatic lesions from 65 patients treated with SBRT at a single institution between January 2021 and November 2023. All lesions received a high biological effective dose of at least 100 Gy, consistent with ablative intent.
The patient population was diverse in tumor origin, with non-small cell lung cancer (35.4%), colorectal cancer (15.4%), and soft tissue sarcoma (10.8%) being the three most common histologies, along with 13 additional cancer types. This histologic heterogeneity reflects real-world clinical practice but also adds complexity to modeling.
Patients were excluded if they had poor performance status (ECOG greater than 2), significant comorbidities, or prior radiotherapy overlapping the treated field. This ensured a cohort with sufficient functional reserve to safely receive high-dose SBRT and adequate follow-up for outcome assessment.
The dataset was split into a 70% training set and 30% test set. All feature selection and model development were performed exclusively on training data to prevent information leakage, with final evaluation on the held-out test set.
Radiomic features were extracted from non-contrast planning CT scans using 3D Slicer with the SlicerRadiomics extension, following Image Biomarker Standardization Initiative (IBSI) guidelines. A total of 851 features per lesion were computed, including morphological, first-order statistical, textural, and wavelet-transform-based features.
To ensure reproducibility, intraclass correlation coefficient (ICC) analysis was performed on 20 randomly selected lesions with repeat segmentations by the same observer at least 4 weeks apart. Only features with ICC of 0.85 or higher were retained, filtering out unstable measurements that could undermine model reliability.
Feature selection for the radiomic model used Elastic Net-regularized Cox regression (Coxnet), which simultaneously performs variable selection and handles correlated features. Hyperparameters (L1 ratio and alpha) were tuned across 300 configurations via 5-fold cross-validation, yielding 1,500 internal model fits.
The final Rad-score was computed as a weighted linear combination of 9 selected radiomic features, including wavelet-based texture and first-order statistics such as kurtosis, skewness, and energy. Lesions were then split into high- and low-risk groups using the training-set median Rad-score as the threshold.
Three separate Cox proportional hazards models were developed: a clinical-only model using features significant in univariable analysis, a radiomic-only model using the Rad-score, and a combined model integrating both. This design allowed direct comparison of the predictive value each data source contributed.
Model performance was assessed using multiple metrics. The concordance index (C-index) measured overall discrimination; area under the receiver operating characteristic curve (AUC) assessed binary classification performance; and time-dependent AUC at 24 months captured discrimination at a clinically meaningful follow-up horizon.
Decision curve analysis was also used to evaluate net clinical benefit across a range of threshold probabilities from 5% to 30%. This method reflects whether acting on model predictions would benefit patients more than treating everyone or no one - a crucial consideration for clinical implementation.
The Akaike information criterion (AIC) compared model parsimony in the training set, rewarding models that achieved better fit without unnecessary complexity. All performance metrics on the test set used 1,000 bootstrap resamples to estimate 95% confidence intervals.
Local recurrence occurred in 12 of 80 lesions (15%) over a median follow-up of 11.8 months. While the overall recurrence rate was relatively low, identifying which specific lesions would recur remains clinically important for guiding treatment intensity and monitoring strategies.
Among clinical variables, soft tissue sarcoma histology and larger metastasis size were the two significant predictors in univariable analysis. In the combined model, soft tissue sarcoma carried a hazard ratio of 7.70, meaning such lesions had nearly 8 times the risk of local recurrence compared to other histologies.
Metastasis size also independently predicted recurrence with a hazard ratio of 1.07 per unit increase, indicating that larger tumors faced incrementally higher recurrence risk - consistent with the established understanding that bigger lesions are harder to ablate completely.
The Rad-score added significant predictive value with a hazard ratio of 4.05 in the combined model. Kaplan-Meier analysis confirmed that patients stratified by Rad-score into high- and low-risk groups had significantly different local recurrence rates on the test set.
All three models demonstrated meaningful discriminative performance. The C-index was 0.75 for the clinical model, 0.74 for the radiomic model, and 0.78 for the combined model on the test set - indicating the combined approach offered the best predictive accuracy.
Conventional AUC values on the test set were 0.74 for clinical, 0.73 for radiomic, and 0.81 for the combined model. At the 24-month time-dependent AUC, performance reached 0.73, 0.71, and 0.80 respectively, confirming that the combined model maintained its advantage at a clinically meaningful time horizon.
The combined model also had the lowest AIC (32.01) compared to clinical (42.71) and radiomic (46.51) models, indicating it provided the best balance of fit and parsimony. The differences in AIC values exceeding 10 points are considered strong evidence of superior model specification.
Decision curve analysis showed that the combined model consistently provided the highest net benefit across threshold probabilities from 5% to 30%, confirming its practical utility for clinical decision-making and risk stratification beyond statistical metrics alone.
This study addresses several gaps in the existing literature on radiomics for SBRT-treated lung oligometastases. Prior studies typically modeled binary treatment response or overall survival rather than time-to-event local recurrence, used less rigorous feature selection, or did not integrate clinical variables systematically.
Previous studies from Cheung et al. and Cilla et al. showed that first-order radiomic features like skewness and surface-to-volume ratio predicted treatment response in lung SBRT, with AUCs ranging from 0.70 to 0.86. However, these studies focused on response rather than recurrence and rarely combined radiomic with clinical predictors.
Fodor et al. specifically examined local recurrence prediction in colorectal cancer lung oligometastases and found that statistical variance of Hounsfield units within the tumor was the most promising feature, though no clinical variables were found to be predictive in that single-histology cohort.
This study advances the field by modeling time-to-event outcomes in a histologically heterogeneous population, enforcing ICC-based reproducibility thresholds, preventing data leakage through strict train-test separation, and demonstrating clinical utility via decision curve analysis - steps that prior work had not consistently implemented together.
If externally validated, model-derived risk estimates could support oncologists in identifying patients likely to experience local failure after SBRT. High-risk patients might be candidates for dose intensification, modified target margins, or closer imaging surveillance following treatment.
Patients predicted to derive limited benefit from standard SBRT might also prompt earlier consideration of alternative local therapies or combined systemic treatment strategies. This represents a shift toward genuinely individualized radiotherapy planning rather than protocol-driven treatment for all patients alike.
Several important limitations must be acknowledged. The study is single-center and retrospective, with only 80 lesions and a median follow-up of under 12 months. The histologically heterogeneous cohort - while clinically representative - limits statistical power and makes generalization uncertain.
The models were not validated on an external dataset, which is a critical step before clinical implementation. The authors also note the need for competing-risk analyses, decision-impact studies, and tools such as a simple online risk calculator with site-specific recalibration to enable practical deployment.
Machine learning models integrating clinical and radiomic features show meaningful potential for predicting local recurrence after SBRT in lung oligometastases. In this cohort, combining both data types improved predictive accuracy over either type alone, supporting the additive value of radiomics when applied rigorously.
The three key predictors identified - soft tissue sarcoma histology, metastasis size, and the Rad-score - are all accessible from routine clinical data and pre-treatment CT scans, meaning they could in principle be integrated into clinical workflows without requiring additional procedures or tests.
Future research priorities include validation in larger, independent, multicenter cohorts; investigation of additional imaging biomarkers such as PET-derived features; and prospective evaluation of model-guided treatment decisions. Standardized and open reporting of models, including model cards and code, would also improve reproducibility and trust in the field.