KEY POINTS
- The study used 50 pediatric proton CSI datasets from the National Cancer Center Korea. An nnU-Net v2 model was trained on 40 patients and independently evaluated on the remaining 10.
- The model automatically generated brain and spine clinical target volumes following an institutional workflow based on SIOPE principles, including the entire cerebrospinal-fluid pathway and pediatric vertebral-body considerations.
- Across the independent evaluation cohort, mean Dice similarity coefficient was 0.9587 ± 0.0305, with mean intersection-over-union 0.9222 ± 0.0553.
- Surface agreement was also strong: mean 95th-percentile Hausdorff distance was 2.49 ± 1.35 mm and mean average symmetric surface distance was 0.79 ± 0.45 mm.
- Brain segmentation was more consistent than spine segmentation. Visual analysis showed the largest discrepancies near inferior boundaries of the brain and spine target volumes, where clinically relevant manual review remains necessary.
- The reported target-delineation workflow fell from approximately eight hours manually to 30 minutes using automated contours plus review and correction. Clinical Likert ratings were predominantly positive to neutral.
- The model generated CTVs rather than final treatment-ready PTVs. Institutional post-processing was still required, including spine expansion and modification around vertebral anatomy and the oesophagus.
- Training came from one institution and only 10 patients formed the test set. Anatomical anomalies, major postoperative distortion, age-dependent body morphology, and external-centre generalizability were not adequately tested.
CLINICAL TAKEAWAY
Automated CSI target delineation could eliminate hours of repetitive contouring while preserving expert review for the anatomically difficult boundaries. The magnitude of the workflow gain is impressive, but multicentre validation is needed before the model can be trusted across the full heterogeneity of pediatric patients.