Application of Radiomics in Predicting Local Control in Postoperative Stereotactic Radiotherapy for Brain Metastasis Patients

Academic Background

Brain Metastases (BMs) are the most common malignant brain tumors, far surpassing primary brain tumors like gliomas in incidence. Recent medical guidelines recommend surgical treatment for patients with symptomatic or larger brain metastases. To improve local control rates, stereotactic radiotherapy (SRT) of the resection cavity is recommended for patients with one or two resected BMs. This method can achieve a local control rate of 70% to 90% within 12 months post-surgery. However, even with adjuvant SRT, the risk of local failure (LF) remains, leading to a demand for pre-treatment radiomic predictive tools to identify patients at high risk of LF.

Main Information of the Study

This study, conducted by Josef A. Buchner et al., was published in the journal “Neuro-Oncology”. The research utilized radiomics and clinical features to develop and externally validate a pre-treatment radiomic-based machine learning model to predict freedom from local failure (FFLF) in patients following brain metastasis surgery and SRT.

Research Methods

Data Collection and Initial Processing

The data for this study was sourced from the multicenter retrospective study “Analysis of Resection Cavity Stereotactic Radiotherapy for Brain Metastases” (Aurora). The training cohort included 253 patients from two centers, while the external test cohort included 99 patients from five centers. Radiomic features were extracted from contrast-enhanced BMs (T1-CE MRI sequences) and surrounding edema (FLAIR sequences). The final model was trained on the entire training cohort and tested on the external test set.

Feature Extraction and Processing

• DICOM images were converted to NIFTI format, and the Brain Extraction Toolkit (BET) was used to extract images containing only the brain.
• Radiomic features were extracted from 3D MRI sequences using the PyRadiomics toolkit.
• NeuroCombat was employed for batch correction to address variations generated by different MRI scanners.
• The Minimum Redundancy Maximum Relevance (mRMR) feature selection framework was applied to select relevant and non-redundant features.

Model Training and Testing

Elastic Net Regression Model (ENR), Random Forest Model (RF), and Extreme Gradient Boosting Model (XGBoost) were used for model training and testing. The optimal batch correction mode and the best feature number were determined via 5-fold cross-validation. The final model was trained on the training set and tested on the multicenter external test set.

Evaluation Metrics

Model performance was quantified using the Concordance Index (CI). For clinical outcomes consideration, decision curve analysis was evaluated at 24 months. The automatically generated segmentations were compared with manual segmentations using the Dice Similarity Coefficient (DSC).

Results and Discussion

Internal Validation and External Testing

In internal validation, the Comb+Pre-Op feature set combined with the ENR learner achieved the highest mean CI of 0.67. In the external test set, the Comb+Pre-Op feature set reached a CI of 0.77, outperforming any clinical feature-only model.

Clinical Application Value

The model significantly differentiated between low-risk and high-risk patient groups (p < 0.001). At 24 months, LF was observed in 9% of the low-risk group and 74% of the high-risk group, showing potential value in patient follow-up and treatment plan adjustment.

Model Performance and Radiation Dose Relationship

The study found that increasing the radiation dose EQD2 (equivalent to a 2 Gray dose) had little impact on local control. In the combined prediction model, prediction performance improved but did not significantly change the model’s clinical value.

Relationship Between Radiomic Features and Tumor Volume

Radiomic features could predict local failure well and were not directly representative of the size or volume of BMs. Additionally, tumor volume showed a low CI in the internal validation of the training cohort but performed well in the external test.

Implications and Recommendations

High-risk patients may benefit from risk-adapted therapy and more frequent follow-ups, such as increased SRT doses, use of systemic drugs that cross the blood-brain barrier, and broader CTV boundaries.

Conclusion

This study developed a machine learning model based on radiomics and clinical features, more effectively predicting the freedom from local failure post brain metastasis surgery. The study validated the feasibility of multicenter datasets, providing a better basis for personalized treatment with the potential to significantly enhance overall management of brain metastasis patients. The model and algorithm are published as an easily accessible web application (https://jbuchner.shinyapps.io/shiny/), facilitating clinicians’ adjustment of follow-up and treatment plans.

This paper has significant innovative implications for the application of combining radiomics with clinical features, potentially providing new perspectives and methods for future patient management and treatment. The study provides a solid data foundation for multicenter clinical trials and demonstrates the potential of radiomic-based predictive models in real-world applications.