3690 - Development of a Fully Automated CTV Segmentation Model for Resection Cavities of Brain Metastases in a Multi-Center Patient Cohort
Presenter(s)
M. Nguyen1, E. Belli1, T. Martins1, C. Zimmer2, B. Meyer3, R. El Shafie4,5, J. Debus6,7, R. Wolff8,9, O. Blanck9,10, K. Eitz1,11, B. Wiestler2, D. Bernhardt1,11, S. E. Combs1,12, and J. C. Peeken1,12; 1Department of Radiation Oncology, Technical University of Munich, Klinikum rechts der Isar, Munich, Germany, 2Department of Diagnostic and Interventional Neuroradiology, Klinikum rechts der Isar, Technical University of Munich, Munich, Germany, 3Department of Neurosurgery, Klinikum rechts der Isar, Technical University of Munich, Munich, Germany, 4Heidelberg University Hospital, Dept. of Radiation Oncology, Heidelberg, Germany, 5Department of Radiation Oncology, University Medical Center Göttingen (UMG), Göttingen, Germany, 6Heidelberg Institute of Radiation Oncology (HIRO), Heidelberg, Germany, 7Department of Radiation Oncology, Heidelberg University Hospital, Heidelberg, Germany, 8Department of Neurosurgery, University Hospital Frankfurt, Frankfurt, Germany, 9Saphir Radiosurgery Center Frankfurt and Northern Germany, Kiel, Germany, 10Department of Radiation Oncology, University Medical Center Schleswig Holstein, Kiel, Germany, 11German Consortium for Translational Cancer Research (DKTK), Partner Site Munich, Munich, Germany, 12Institute of Radiation Medicine (IRM), Department of Radiation Sciences (DRS), Helmholtz Center Munich, Munich, Germany
Purpose/Objective(s): Clinical Target Volume (CTV) segmentation is a time-intensive task in radiation treatment planning and is prone to interrater variability. Automated segmentation of resection cavities (RCs) and their CTVs remains challenging due to postoperative changes and the dura mater’s intricate structure. This study presents an automated workflow to segment RCs and to subsequently interactively define the CTV by applying intra-parenchymal and dural margins.
Materials/Methods: Data was collected within a multicenter study, which included postoperative T1c-MR images from 246 patients across three centers. Initially, automated segmentation of the RCs was established using deep learning. A nnUNet v2 model was trained on 180 cases from Center 1 using 5-fold cross-validation. The model was then tested on an external dataset, which included 24 cases from Center 2 and 42 cases from Center 3. To define the CTV, a nnUNet v2 model for automated segmentation of the dura mater was trained using 24 manually annotated cases, including the falx cerebri and tentorium cerebelli. According to an international guideline, python-based scripts were developed to add intra-parenchymal safety margins to the RCs and adaptive dural margins to the dura mater [1]. To ensure that the dura mater expands only along the adjacent layer, a script was developed to precisely identify its inner surface facing the RC and to fully encompass its entire thickness while expanding it by a specified margin. Additionally, by delineating the dura mater, the script ensured that intra-parenchymal expansion remained confined within the brain tissue.
Results: The nnUNet model for RC segmentation achieved a median Dice Similarity Coefficient (DSC) of 0.90 ± 0.04 (MAD) on the validation set and 0.91 ± 0.04 (MAD) on the independent test set. The interrater variability was estimated at 0.85 ± 0.06 (MAD) in the test set, based on center-specific segmentations and newly created manual annotations created for this study. Pearson correlation analysis showed a dependency between volume size and DSC (overall score of 0.33, p = 0.008), particularly for volumes under 6 cm³. For dura segmentation, the model achieved a median validation DSC of 0.81 ± 0.02 (MAD). A comparison between the automatically generated CTVs and the manually generated CTVs demonstrated an agreement with a median test DSC of 0.93 ± 0.03 (MAD).
Conclusion: Our proposed model for RC segmentation demonstrated robust performance, achieving a DSC score comparable to interrater variability. Expanding the training dataset is expected to further improve performance. The clinical applicability of our interactive automatic CTV segmentation model requires further evaluation.