A high-stakes clinical problem. Small cell lung carcinoma (SCLC) is an aggressive neuroendocrine malignancy with a high rate of brain metastasis. Between 10% and 20% of SCLC patients already have brain metastases at the time of initial diagnosis, making intracranial spread a major contributor to mortality in this population.
Standard treatment and its limitations. Whole-brain radiotherapy (WBRT) is the established first-line treatment for SCLC patients with diffuse brain metastases. However, individual patients respond very differently to this therapy, and predicting who will benefit remains difficult with conventional clinical factors alone.
Prior prediction efforts. Earlier studies used markers like gender, age, extracranial metastasis status, and TNM staging to build predictive models, but these clinical variables capture only a limited picture of tumor biology and heterogeneity. More data-rich approaches were needed to improve prediction accuracy.
The promise of imaging-based biomarkers. Radiomics, which extracts high-dimensional quantitative features from medical images, has shown promise for predicting treatment response in multiple cancers. Dosiomics, a related approach that analyzes the spatial distribution of the radiation dose, adds complementary information not captured by traditional dose-volume histograms.
Core goal. This study aimed to identify predictors of WBRT response in SCLC brain metastasis patients and to build accurate machine learning models by combining clinical data with both radiomics and dosiomics features in a multi-omics framework.
Patient cohort. The study retrospectively analyzed 144 SCLC patients with brain metastases treated at Yunnan Cancer Hospital between January 2020 and June 2024. Patients were classified as responders (complete or partial response, n=74) or non-responders (stable or progressive disease, n=70) based on their first head MRI follow-up one to three months after treatment.
Seven models compared. Seven classification models were constructed ranging from single-omics models using only clinical factors, radiomics, or dosiomics features, to combination models pairing clinical with imaging data, up to a hybrid model (HFM) integrating all three data types. The HFM was designated the primary model of interest.
External validation. To assess generalizability beyond the training cohort, patients treated prospectively from January to June 2025 were collected as an independent external validation set, providing a real-world test of the model's performance.
Two imaging data sources. Features were extracted from two complementary image types: pre-treatment CT scans of the brain and three-dimensional radiation dose distribution images produced by the treatment planning system (TPS). Both sets were processed using 3D Slicer software (version 5.6.2).
Standardized preprocessing. All DICOM-formatted CT scans and dose images were resampled to a uniform voxel size of 1x1x1 mm3 before feature extraction. The region of interest was the entire brain clinical target volume (CTV) used for whole-brain irradiation, ensuring consistent delineation across patients.
Large feature pool. From each image set, 107 original features were extracted including shape features, first-order statistics, and texture features. These were then filtered through wavelet transforms to generate additional feature sets, resulting in a total of 851 omics features from both CT and dose images combined.
Radiotherapy delivery parameters. All patients received VMAT (volumetric modulated arc therapy) using 6-MV photon beams. Treatment planning used either Pinnacle or Monaco systems, with dose calculated by the Monte Carlo algorithm at a 0.3 cm grid resolution, normalized so that 95% of the planning target volume received the prescribed dose.
LASSO regression for dimensionality reduction. Given the large number of extracted features relative to the patient sample size, the study used a two-stage selection process: first, univariate analysis filtered for features significantly associated with treatment response, and then LASSO (Least Absolute Shrinkage and Selection Operator) regression was applied to identify the most relevant predictors while shrinking non-informative coefficients to zero.
Seven features selected. This process identified four radiomics features (original GLCM, wavelet-LLH first-order, wavelet-HLL GLSZM, wavelet-LLL first-order) and three dosiomics features (original shape, wavelet-HHL GLSZM, wavelet-HHH GLSZM) as significant predictors of WBRT response.
Logistic regression modeling. A logistic regression algorithm was used to build all seven models. Model performance was evaluated using AUC, accuracy, sensitivity, specificity, false positive rate, and F1 score. For the best-performing hybrid model, calibration curves and decision curve analysis (DCA) were also generated.
Validation strategy. Internal validation used bootstrap resampling with 1000 iterations to estimate model stability. A nomogram was constructed from the hybrid model's features to provide individualized visual predictions for clinical use. An independent prospective external dataset was used for final generalizability testing.
Hybrid model performed best. Among all seven models, the hybrid model (HFM) combining clinical factors, radiomics, and dosiomics achieved the highest AUC of 0.792, with accuracy of 72.2%, sensitivity of 77.0%, and specificity of 71.4%. This represented a meaningful improvement over each single-omics model.
Dosiomics outperformed radiomics alone. A notable finding was that the dosiomics model (DF, AUC 0.699) performed better than the radiomics model (RF, AUC 0.682) when used in isolation. This was attributed to dosiomics capturing subtle spatial differences in dose delivery that predict whether underdosed regions within the target volume would drive treatment failure.
Independent clinical predictors identified. Multivariate logistic regression confirmed that concurrent chemoradiotherapy (CCRT), conformal boost radiotherapy (CBRT), dosiomics score, and radiomics score were each independent predictors of WBRT response. Immunotherapy showed significance in univariate analysis but was not retained after multivariate adjustment.
Strong partial AUC in high-specificity range. The hybrid model achieved a partial AUC of 0.038 when the false positive rate was constrained to 10% or below, indicating that the model maintains strong discriminative power in clinical scenarios where high specificity is critical for avoiding unnecessary treatment escalation.
Excellent model calibration. The calibration curve for the hybrid model showed a Brier score of 0.011, a slope of 1.02, and an intercept of -0.031, indicating that predicted probabilities closely matched observed response rates. This level of calibration is important for translating model outputs into actionable clinical probability estimates.
Nomogram for individualized prediction. A nomogram was constructed integrating the dosiomics score, radiomics score, CCRT status, and CBRT status. This tool allows clinicians to input patient-specific values and obtain a visual probability estimate for WBRT response, supporting individualized treatment decision-making.
External validation confirmed generalizability. When applied to the prospectively collected external validation cohort, the nomogram achieved an AUC of 0.711 (95% CI 0.487-0.934). While the confidence interval was wide due to the small validation sample, the result supports the model's ability to generalize beyond the training institution.
Internal bootstrap validation. With 1000 bootstrap iterations, the model achieved a mean AUC of 0.792 (95% CI 0.708-0.852), a Youden index of 0.485, sensitivity of 77.0%, and specificity of 71.4%. Bootstrap validation provides a more honest estimate of model performance by accounting for overfitting.
Spatial dose heterogeneity matters. Traditional dose-volume histograms only report aggregate dose statistics and cannot capture where within the target volume the dose is lowest. Dosiomics features encode the three-dimensional spatial distribution of dose, capturing subtle local underdosing patterns that correlate with treatment failure even when the global dose prescription appears met.
An unexpected finding on dosiomics superiority. Unlike some prior studies where radiomics outperformed dosiomics, this study found the opposite for WBRT in SCLC. The authors attribute this to the fact that for whole-brain irradiation, the radiomics and dosiomics feature sets are drawn from the same anatomical region (the whole brain CTV), leading to high feature overlap (Pearson correlation 0.236, p=0.004) that partially masks the radiomic signal.
Concurrent treatments as independent predictors. The significance of CCRT and CBRT as independent predictors aligns with prior literature showing that concurrent chemotherapy sensitizes brain metastases to radiation, and that local dose escalation via conformal boost improves local control. These clinical factors complement the imaging-based predictors in the hybrid model.
Survival analysis findings. Kaplan-Meier analysis showed that patients in the radiation response group tended to have longer progression-free survival, but no significant difference in overall survival between responders and non-responders was observed. The authors caution that this finding may reflect the limited sample size rather than a true biological equivalence in survival outcomes.
A new biomarker paradigm. This study establishes dosiomics features extracted from radiotherapy dose images as a novel class of biomarkers for predicting treatment response in SCLC brain metastasis. The approach is entirely non-invasive, using data already generated as part of routine radiotherapy planning.
Practical clinical tool. The nomogram integrating clinical factors, CT radiomics, and dose-derived dosiomics provides a ready-to-use decision support tool. By predicting who is likely to respond to WBRT before treatment, oncologists can potentially individualize therapy -- escalating dose or adding concurrent chemotherapy for predicted non-responders while protecting responders from unnecessary treatment intensification.
Acknowledged limitations. The study is single-center and retrospective, limiting immediate generalizability. Test-retest reliability of dosiomics features was not formally assessed. Different linear accelerators, treatment planning systems, and dose calculation algorithms may affect dosiomics feature stability across institutions. The external validation cohort was small.
Future work. The authors plan to pool data across multiple institutions to build a more comprehensive and universally applicable nomogram. Expanding to WBRT patients with other primary cancers such as breast cancer and NSCLC will allow testing whether the multi-omics framework generalizes across tumor histologies.