AURORA trial : A Multicenter Analysis of Stereotactic Radiotherapy to the Resection Cavity of Brain Metastases
Study PI:
PD Dr. Denise Bernhardt (Part 1)
Technical University of Munich (TUM)
denise.bernhardt@mri.tum
PD Dr. Jan Peeken (Part 2)
Technical University of Munich (TUM)
jan.peeken@tum.de
Prof. Dr. med. Stephanie E. Combs
Technical University of Munich (TUM)
stephanie.combs@tum.de
Study design:
● Multi-center, international retrospective study
Study objective:
● Part 1: clinical evaluations regarding prognosis, fractionation, toxicity, OS, PFS, surgery, GPA, RPA scoring
● Part 2: radiomics and deep learning analysis for metastases characterizration, prognostic assessment and segmentation.
Project status:
● Data collection and transfer completed at all centers
● Statistical analysis ongoing (Part 1)
● Development of radiomics and deep learning models ongoing (Paart 2)
GitHub:
● MRI-based UNet segmentation model for brain metastases: http://github.com/neuronflow/AURORA
Major findings and results:
Buchner, J.A., Kofler, F., Mayinger, M., Christ, S.M., Brunner, T.B., Wittig, A., Menze, B., Zimmer, C., Meyer, B., Guckenberger, M., Andratschke, N., el Shafie, R.A., Debus, J., Rogers, S., Riesterer, O., Schulze, K., Feldmann, H.J., Blanck, O., Zamboglou, C., Ferentinos, K., Bilger-Zähringer, A., Grosu, A.L., Wolff, R., Piraud, M., Eitz, K.A., Combs, S.E., Bernhardt, D., Rueckert, D., Wiestler, B., Peeken, J.C. (2024). Radiomics-based prediction of local control in patients with brain metastases following postoperative stereotactic radiotherapy. Neuro-Oncology, 26(9), 1638–1650. https://doi.org/10.1093/neuonc/noae098
- Data: Patient data from the AURORA retrospective study were used. A total of 253 patients from 2 centers formed the training cohort; 99 patients from 5 additional centers were included as an external test cohort. All patients had surgical resection of a large or symptomatic brain metastasis (BM).
- Method: Radiomic features were extracted from the contrast-enhancing portion of the BM (T1-CE MRI) and the surrounding edema (T2-FLAIR) pre-therapy. Multiple models combining radiomic and clinical variables were trained via elastic net regression, with optimal parameters determined by a 5-fold cross-validation.
- Key Findings:
- The best performing model combined radiomic and clinical features (concordance index [CI]: 0.77) and outperformed any clinical-only model (best CI: 0.70).
- Patients were effectively stratified by local failure (LF) risk (p < .001), with an incremental net clinical benefit.
- At 24 months, LF rates were 9% in the low-risk group versus 74% in the high-risk group
- Conclusion and Implications: Integrating clinical and radiomic features into a predictive model yields better estimation of LF risk than clinical variables alone. High-risk patients may benefit from more frequent follow-up or intensified adjuvant therapy.
Buchner, J. A., Kofler, F., Etzel, L., Mayinger, M., Christ, S. M., Brunner, T. B., Wittig, A., Menze, B., Zimmer, C., Meyer, B., Guckenberger, M., Andratschke, N., El, R. A., Debus, J., Rogers, S., Riesterer, O., Schulze, K., Feldmann, H. J., Blanck, O., … Peeken, J. C. (2023). Development and external validation of an MRI-based neural network for brain metastasis segmentation in the AURORA multicenter study. Radiotherapy and Oncology, 178, 109425. https://doi.org/10.1016/j.radonc.2022.11.014
- Data:348 patients from six centers with at least one brain metastasis originating from various primary cancers. All received MRI scans (T1 ± contrast, T2, FLAIR).
- Method:A 3D U-Net was trained on a cohort of 260 patients from two centers to segment both the GTV (gross tumor volume) and the surrounding FLAIR hyperintense region. Manual reference segmentations for tumors and edema were created. Data augmentation strategies were employed during training. An external multicenter test cohort (88 patients from four additional centers) was used for validation.
- Key Findings for Tumor Segmentation:
- The U-Net achieved a mean Dice similarity coefficient (DSC) of 0.92 ± 0.08 across all patients and 0.89 ± 0.11 on an individual metastasis basis.
- Data augmentation significantly improved segmentation performance.
- Metastasis detection was effective, with a mean F1-score of 0.93 ± 0.16.
- Performance was stable across different centers (p = 0.3), and metastasis volume had no correlation with DSC (Pearson r = 0.07).
- Conclusion and Clinical Relevance: Automated neural network–based segmentation of brain metastases is both reliable and consistent across centers. This approach may streamline radiotherapy planning by providing objective, consistent GTV definitions.
Buchner, J. A., Peeken, J. C., Etzel, L., Ezhov, I., Mayinger, M., Christ, S. M., Brunner, T. B., Wittig, A., Menze, B., Zimmer, C., Meyer, B., Guckenberger, M., Andratschke, N., el Shafie, R. A., Debus, J., Rogers, S., Riesterer, O., Schulze, K., Feldmann, H. J., … Kofler, F. (2023). Identifying core MRI sequences for reliable automatic brain metastasis segmentation. Radiotherapy and Oncology, 188, 109901. https://doi.org/10.1016/j.radonc.2023.109901.
- Data:339 patients with brain metastases from seven centers were included.
- Method:Automated brain metastasis (BM) and peritumoral edema segmentation was conducted using a 3D U-Net, trained on four different MRI sequences (T1, T1 with contrast-enhancement (T1-CE), T2, and T2-FLAIR) in various combinations.
- Key Finding for Tumor Segmentation:T1-CE alone provided the best BM segmentation performance (Dice = 0.96). Models without T1-CE had significantly lower accuracy (T1-only: Dice = 0.70, T2-FLAIR-only: Dice = 0.73).
- Key Finding for Edema Segmentation:The combination of T1-CE and T2-FLAIR achieved the highest edema segmentation accuracy (Dice = 0.93). Omitting either T1-CE or T2-FLAIR resulted in reduced performance (Dice = 0.81–0.89).
- Conclusion and Implications:T1-CE alone is sufficient for optimal BM segmentation. However, for edema segmentation, T1-CE and T2-FLAIR together are essential. These insights may streamline clinical protocols by omitting unnecessary sequences and reducing scanning times while maintaining accurate, automated target definition.