S2788
RTT - RTT contouring, target definition, and treatment planning
ESTRO 2026
overview. https://doi.org/10.2217/cns.14.512. Ohira S, et al. (2018) HyperArc VMAT planning for single and multiple brain metastases stereotactic radiosurgery. https://doi.org/10.1186/s13014-017-0948-z3. Garsa A, et al. (2021) Radiation Therapy for Brain Metastases: A Systematic Review.https://doi.org/10.1016/j.prro.2021.04.002 Keywords: Brain Metastases, SRS, HyperArc VMAT Digital Poster 2950 Generalization of automated head organ segmentation to various MR image sequences Eszter Szabó 1 , Ádám Kékesi 2 , Eszter K. Kiss 2 , Bernadett Kolozsvári 2 , Zsófia Karancsi 2 , Eszter Ruff 2 , Botond Sipos 2 , Barbara Gregus 2 , Valter Gila 1 , István Megyeri 1 , Krisztián Koós 1 , Borbála Deák-Karancsi 2 , László Ruskó 2 1 Science and Technology, GE HealthCare, Szeged, Hungary. 2 Science and Technology, GE HealthCare, Budapest, Hungary Purpose/Objective: Deep-learning segmentation is routinely used in radiation therapy for both CT and MR images. In MR modality various sequences provide superior soft- tissue information for organ delineation. The variety of sequences used in practice raises the need for developing sequence-agnostic models. The goal of this work was to train and test models for head organ segmentation using frequent MR sequences. Material/Methods: The training data involved 2 datasets: T1(C) including 50 (3D BRAVO) scans (30/20 w/wo contrast), and T2 including 50 (2D FSE) scans. The test data involved 3 datasets: T1(C) including 10 (3D BRAVO) scans (6/4 w/wo contrast), T1 including 40 (3D MPRAGE) scans (wo contrast), and T2 including 33 (2D FSE) scans. In all datasets 14 structures (body + 13 organs) were manually contoured.Using the nnU-NET framework [1] three segmentation models were trained: using the 50 T1(C) scans (T1-specific), using the 50 T2 scans (T2- specific), and using 25-25 T1 and T2 scans (hybrid). The models were trained until convergence with default settings derived by the nnUNET framework. The output of the model was a multi-label map, where labels 1-14 represent the structures.Each segmentation model was evaluated without any post- processing, only axial hole filling was applied to structures which involve another one. The T1-specific model was tested on T1(C) and T1 data, the T2-specific model was tested on T2 data, while the hybrid model was tested on all test data. DICE score was used to measure the similarity between auto- and manual contours. The mean accuracy was computed for each organ and for the whole test set in all experiments.
Two cohorts were retrospectively analysed: SBM (n=60; SRS VMAT=25, HyperArc=35) and MBM (n=22; 11 p/technique). PCI, SI, HI, and GI were calculated for each plan. Data normality was assessed using the Shapiro–Wilk test. Comparisons between groups were conducted using independent-samples t-tests or Mann–Whitney U tests, as appropriate based on data distribution. Results: No significant differences were observed for PCI or SI in either cohort: for SBM, PCI was 0.924±0.034 (VMAT) versus 0.935±0.046 (HyperArc) and SI 0.951±0.063 vs 0.942±0.037; for MBM, PCI was 0.88±0.06 vs 0.91±0.03 and SI 0.94±0.05 vs 0.92±0.03 (all p>0.12). In contrast, HI and GI demonstrated consistent differences between techniques. For SBM, HI averaged 0.132±0.052 in SRS VMAT vs 0.187±0.049 in HyperArc (U=180.0; p<0.001), while GI was 3.36±0.99 vs 2.81±0.57, respectively (U=219.5; p=0.001). For MBM, SRS VMAT showed lower HI (0.11±0.03) compared to HyperArc (0.20±0.04; t(20)= − 4.883; p<0.001), whereas GI favoured HyperArc (3.02±0.55) over SRS VMAT (4.86±1.52; t(11.827)=3.440; p=0.005). These findings indicate that SRS VMAT consistently produced more homogeneous plans, whereas HyperArc achieved steeper dose gradients, reflecting fundamental differences in dose distribution between the two techniques.
Conclusion: Both techniques demonstrated similar conformity and selectivity (PCI and SI), indicating equivalent target coverage accuracy. However, a consistent and clinically relevant trade-off was observed between homogeneity and dose gradient behavior. SRS VMAT achieved more uniform dose distributions, reflected by lower HI values, which may be advantageous for lesions near sensitive structures. Conversely, HyperArc produced steeper dose gradients (lower GI), optimizing healthy brain tissue sparing, particularly in multiple metastases. These results emphasize the importance of individualized planning strategy selection, balancing homogeneity and gradient performance based on target complexity, location, and clinical objectives. References: 1. Bertolini F, et al. (2015) Brain metastases: an
Made with FlippingBook - Share PDF online