S1608
Physics - Autosegmentation
ESTRO 2026
challenging due to their small size, limited soft tissue contrast, and scarcity of high-quality labelled datasets. Enabling reliable CT-based segmentation could unlock large-scale retrospective dose-response analyses where MR images have not been preserved. This study evaluated whether a deep learning model can accurately segment brain substructures from CT alone when trained with MR-derived labels in paediatric and young adult medulloblastoma patients. Material/Methods Paired planning CT-MR datasets from 43 medulloblastoma patients (age 2-24 years) treated with proton craniospinal irradiation were used. T1- weighted MR images were automatically segmented into eight representative brain substructures – the brainstem, thalami, hippocampi, caudate nuclei, lateral ventricles, cerebellum, white matter, and cerebral cortex - using FreeSurfer v7.4.1[1]. Labels were copied to CT, after rigid registration, to create CT-domain training data. Left and right substructures were merged to create single bilateral labels for each region before training and evaluation. A 3D nnU-Net ([2]) was trained exclusively on CT data, and evaluated using five-fold cross-validation. Performance was assessed using mean distance to agreement (mDTA) and 95 th percentile Hausdorff distance (HD95), Figure 1. Relationships between segmentation accuracy, structure volume, and CT intensity variability were examined.
mm). Geometric accuracy was lowest in the brainstem and cerebellum, regions commonly affected by tumour presence, surgical resection, and shunt artefacts (mDTA 7.4±3.8 mm; HD95 21±9 mm). These posterior fossa regions also exhibited greater intensity heterogeneity, consistent with disrupted anatomy and degraded CT contrast. No trends between structure volume and segmentation accuracy were observed.
Conclusion CT-based nnU-Net segmentation demonstrates encouraging accuracy for several paediatric brain substructures but remains challenging in tumour- affected posterior fossa regions typical of medulloblastoma patients. These findings underline the challenges of CT-based contouring in this population while demonstrating the potential of MR- derived labels to support model training. This work provides a foundation for further optimisation and the development of automated, large-scale tools to explore dose–response and late effects in paediatric brain tumour cohorts. References [1] Fischl B. FreeSurfer. Neuroimage. 2012;62(2):774– 781. doi:10.1016/j.neuroimage.2012.01.021 [2] Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier- Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18(2):203–211. doi:10.1038/s41592-020- 01008-z Keywords paediatric, CT brain segmentation, medulloblastoma
Results Segmentation accuracy varied across brain regions (Figure 2). The cerebral cortex, white matter, caudate, and hippocampi showed the best geometric agreement (mDTA 0.7±0.4 mm; HD95 2.4±0.9 mm). Intermediate performance was seen for the thalami and ventricles (mDTA 3.4±2.5 mm; HD95 10.5±8.2
Made with FlippingBook - Share PDF online