Accurate survival prediction for lung cancer patients receiving radiotherapy remains difficult. Non-small cell lung cancer patients treated with radiation therapy show highly variable outcomes that are not fully captured by existing clinical staging systems. Identifying which patients will respond well versus poorly to treatment could guide more personalized follow-up and combination therapy decisions.
Imaging biomarkers offer a potential path to better prognosis without invasive procedures. CT scans are routinely acquired before and after radiotherapy, creating a natural source of longitudinal data about tumor changes over time. Radiomics approaches that extract quantitative features from these images have shown promise as prognostic tools, but classical radiomics requires manual feature engineering and is sensitive to acquisition variability.
Foundation models offer a fundamentally different approach to medical image analysis. Pre-trained on large, diverse image datasets, foundation models learn generalizable representations without task-specific supervision. The Foundation Model for Cancer Imaging Biomarkers (FMCIB) uses a 3D ResNet50 architecture pre-trained on more than 11,467 lesion images from diverse cancer types, enabling rich feature extraction from CT images that does not depend on local dataset size.
Temporal changes between pre- and post-treatment scans may encode treatment response. Rather than analyzing a single CT scan, this study tested whether changes between baseline and post-treatment scans could capture biological information about tumor response to radiation that predicts long-term survival outcomes beyond what either scan alone can provide.
102 NSCLC patients treated at Dana-Farber/Brigham Cancer Center formed the study cohort. All patients received radiation therapy between 2002 and 2014, yielding 204 CT scans in total (one pre-treatment and one post-treatment per patient). Median follow-up was 11.4 months, with 1-year and 2-year overall survival as the primary endpoints alongside concordance index (C-index) for overall discriminative performance.
The FMCIB model extracted 4096-dimensional feature vectors from CT patch regions. For each patient, a 50x50x50 mm cube centered on the annotated tumor seed point was extracted from both the pre-treatment and post-treatment scans. The FMCIB 3D ResNet50 processed each patch independently, producing a high-dimensional embedding that captures the spatial texture, density, and structural characteristics of the tumor region.
Four feature representation strategies were systematically compared. The study tested: pre-treatment features alone, post-treatment features alone, element-wise subtraction of pre from post features (a 4096-dimensional difference vector), and the Euclidean distance (L2 norm) of the subtracted vector as a single scalar summary of change. These were evaluated both alone and combined with a clinical model incorporating age, sex, performance status, TNM stage, histology, and radiotherapy dose.
Cox proportional hazards regression linked features to survival outcomes. The high-dimensional features were reduced using PCA before Cox regression for the full feature approaches, while the Euclidean distance summary was entered directly as a scalar predictor. Models were evaluated by leave-one-out cross-validation given the limited sample size. Feature importances were assessed using permutation-based methods to identify which feature dimensions most contributed to survival prediction.
Subtracted features significantly outperformed all other feature representations. The element-wise difference between post- and pre-treatment FMCIB features achieved a C-index of 0.74, a 1-year AUC of 0.93, and a 2-year AUC of 0.91 (p equal to 0.006). This substantially exceeded the pre-treatment-only model (C-index 0.55), the post-treatment-only model (C-index 0.64), and the clinical-only model (C-index 0.46).
The Euclidean distance scalar provided modest but meaningful improvement over clinical features alone. When the L2 norm of the subtracted feature vector was added to the clinical model, the combined model achieved a C-index of 0.53, compared to 0.46 for clinical features alone. While this improvement was smaller than the full subtracted feature approach, the Euclidean distance offers a much simpler representation that could be integrated into clinical workflows more easily.
Neither pre-treatment nor post-treatment features alone matched the temporal difference approach. The superiority of the subtracted feature approach over both single time-point models confirms that temporal change information is the key prognostic signal rather than either baseline tumor characteristics or post-treatment state independently. This supports the conceptual importance of treatment response as a survival determinant beyond static tumor measurements.
The subtracted feature model was statistically significant while single time-point models were not. Only the subtracted feature approach reached statistical significance (p equal to 0.006), while pre-treatment features (p equal to 0.12), post-treatment features (p equal to 0.07), and Euclidean distance alone (p equal to 0.12) did not individually achieve significance in this relatively small cohort.
The first PCA component of subtracted features correlated strongly with tumor size. Pearson correlation between the first principal component of the subtracted FMCIB features and maximum tumor diameter was r equal to 0.57 (p less than 1x10-9). This finding confirms that the model's most prominent learned dimension captures the physical change in tumor size after treatment, a clinically meaningful and expected prognostic indicator.
Feature importance analysis identified which dimensions drove survival predictions. Permutation-based importance testing of the top PCA components showed that only a subset of the reduced feature dimensions contributed meaningfully to Cox model performance. This suggests that while the original FMCIB representation is high-dimensional, the survival-relevant signal concentrates in a limited number of learned imaging dimensions related to treatment response.
The correlation with tumor size validates the biological plausibility of the approach. Classical radiomics studies have shown that tumor volume response to radiation is associated with overall survival. The finding that the foundation model's primary learned component aligns with this established prognostic factor indicates that FMCIB features are capturing true biological information rather than statistical artifacts of the small dataset.
Subtracted features capture response beyond tumor shrinkage alone. While the first component correlates with tumor size change, higher-order components encode additional texture and structural information about treatment response not captured by volume measurements alone. This suggests that the full 4096-dimensional representation may contain complementary prognostic information across multiple biological dimensions of tumor response.
Foundation models eliminate the need for manual feature engineering. Traditional radiomics pipelines require expert selection and validation of handcrafted features such as GLCM texture metrics, shape descriptors, and intensity histograms. FMCIB extracts a single fixed 4096-dimensional representation automatically, reducing pipeline complexity and removing subjective decisions about which features to include. This standardization could improve reproducibility across centers.
The single-center retrospective design is the study's primary limitation. With 102 patients from one institution and a median follow-up of only 11.4 months, the findings cannot yet be generalized to other clinical settings. External validation on independent cohorts from different institutions, treatment protocols, and patient demographics is required before these results can influence clinical practice.
The model analyzes only the primary tumor region, potentially missing systemic response information. Both pre- and post-treatment patches were extracted from the same annotated seed point in the primary tumor. Information about lymph node response, pleural changes, or systemic effects of treatment visible elsewhere on the CT scan was not incorporated. Expanding the spatial scope of the analysis could capture additional prognostic signals.
Small sample size limits statistical power for subgroup analyses. With 102 patients, the study cannot reliably evaluate whether the approach performs differently across histological subtypes (adenocarcinoma versus squamous cell), treatment modalities (conventional fractionation versus SBRT), or disease stages. Larger prospective datasets would enable these clinically important subgroup analyses.
Foundation model-based analysis of temporal CT changes significantly outperforms single time-point approaches. This study demonstrates that pairing pre- and post-treatment CT scans with element-wise feature subtraction from a pre-trained model captures prognostic information that neither scan alone provides. The approach requires no task-specific model training, relying entirely on a generalizable foundation model pre-trained on a large external dataset.
The method has practical advantages for real-world implementation. Because FMCIB requires only a small image patch around the tumor seed point and applies a fixed pre-trained network without any additional training on local data, the approach is computationally efficient and does not require large institutional datasets to implement. This may make it particularly useful at centers with limited patient volumes or computing resources.
Identifying patients at risk of poor survival could guide treatment intensification decisions. Patients predicted to have poor survival outcomes based on post-treatment CT changes could be candidates for more aggressive follow-up imaging, earlier consideration of salvage therapy, or enrollment in trials of treatment intensification. The ability to make these predictions from routine CT scans without additional testing is a key clinical advantage.
Prospective validation and expansion to other cancer types are the natural next steps. The FMCIB model was pre-trained on diverse cancer imaging data, suggesting this temporal subtraction approach may generalize beyond lung cancer to other cancer types where paired pre- and post-treatment CT scans are routinely acquired. Prospective multicenter trials will be needed to confirm clinical utility and establish appropriate thresholds for clinical decision-making.