Deep learning segmentation of the cerebral ventricular system and brainstem for pediatric radiotherapy planning.

BACKGROUND AND PURPOSE: Segmentation of the cerebral ventricular system (VS) is central to radiotherapy, both for whole-ventricular irradiation in intracranial germ cell tumors (GCTs) and protection of the periventricular region. Complex, variable VS anatomy makes manual delineation challenging and hampers auto-segmentation model development. We reported the evaluation and workflow integration of a deep-learning (DL) model for VS and brainstem segmentation on pediatric magnetic resonance imaging (MRI).

MATERIALS AND

METHODS

An nnU-Net was trained on MRIs from 76 unique brain tumor cases (n = 73/71 for T1/T2). The evaluation used two independent retrospective cohorts: (i) GCTs evaluated on planning computed tomography (CT) images after MRI-to-CT registration mapping (n = 13/14), and (ii) brain tumors evaluated on MRI (n = 14). Performance was assessed using the Dice similarity coefficient (DSC) and Hausdorff distances (HD) against manual contours. Prospective deployment in routine clinical planning assessed usability and workflow integration.

RESULTS

In the GCT-CT cohort, brainstem segmentations achieved median DSCs of >0.85 and HDs95 of ≤5.0/6.4 mm, while VS segmentations achieved DSCs of >0.70 and HDs95 of ≤5.3/4.3 mm, after rigid/deformable mapping to planning CT. In the MR cohort, performance on T1-/T2-weighted MRI reached DSCs of ≥0.90 for both structures, and HDs95 of ≤3.0/3.4 mm and ≤ 1.3/2.2 mm for the brainstem and VS, respectively. The DL-segmentation workflow reduced workflow time by >50% compared with manual contouring.

CONCLUSIONS

Lower performance in the GCT-CT cohort likely reflects noisier CT-based references, interobserver variability, and uncorrected geometric/anatomical mismatch. The model's accuracy supports radiotherapy planning in pediatric brain tumors through risk-structure delineation and target-volume segmentation in GCTs.