KEY POINTS
- The nnU-Net models were trained on T1- and T2-weighted MRI from 76 pediatric brain-tumor patients ≤18 years, covering germ-cell tumors, medulloblastoma, ependymoma, glioma, AT/RT and other CNS entities.
- Two independent 14-patient validation cohorts were used: one compared predictions directly with expert contours on MRI, while the germ-cell-tumor cohort assessed the full workflow after mapping MRI-generated contours onto the planning CT.
- On direct MRI evaluation, both ventricular-system and brainstem segmentation achieved median Dice scores ≥0.90. Brainstem HD95 was ≤3.0/3.4 mm on T1/T2 MRI, while ventricular-system HD95 was ≤1.3/2.2 mm.
- Performance fell after MRI-to-CT registration, reflecting the additional uncertainty of the clinical workflow. In the GCT planning-CT cohort, brainstem Dice remained >0.85, whereas ventricular-system Dice was >0.70; HD95 remained approximately ≤5–6 mm.
- Figure 3 illustrates where manual corrections were still needed after mapping, including the cranial brainstem border, medulla, prepontine cistern and inferior fourth-ventricle region—important examples of why high global Dice alone does not eliminate physician review.
- In two prospective CNS germ-cell-tumor cases, clinicians rated the generated ventricular and brainstem structures 4–5/5 on MRI and 4/5 after mapping to planning CT. Model inference itself required <1 minute.
- Radiation oncologists estimated that the complete workflow—including segmentation, registration, mapping and manual correction—reduced contouring time by >50% compared with fully manual delineation.
CLINICAL TAKEAWAY
Automated ventricular and brainstem segmentation looks clinically useful in pediatric CNS planning, particularly for whole-ventricular irradiation and evaluation of the radiosensitive periventricular region. The main remaining weakness is not MRI segmentation itself but the registration and mapping step into CT-based workflows, so physician review remains necessary.