KEY POINTS
- The study developed an nnU-Net model to segment 25 organs of interest defined in the EPTN international neurological contouring atlas. The dataset contained 74 adults with primary intracranial tumours, mainly gliomas and meningiomas, including patients with surgical cavities and hydrocephalus.
- Each patient had a planning CT, contrast-enhanced T1-weighted MRI, T2 FLAIR MRI, and reference contours. Five configurations were compared: T1 alone, T1 plus T2 FLAIR, all three modalities, T1 plus CT, and a hybrid model combining separate T1- and CT-based predictions.
- Five-fold cross-validation used 59 patients for training and 14–15 for evaluation in each fold. Every patient was evaluated once on unseen data, producing 74 independent predictions, but no separate external test cohort was available.
- The best-performing configuration combined T1-weighted MRI and CT, achieving a median Dice coefficient of 0.80, surface Dice of 0.84, added path length of 647 mm, and 95th-percentile Hausdorff distance of 2.24 mm across all structures and patients.
- Adding CT increased median Dice by more than 0.10 for several ocular structures, including the corneas, lenses, maculas, and foveas. Adding T2 FLAIR to T1-weighted MRI produced negligible improvement, with median Dice differences below 0.015.
- The T1-plus-CT model outperformed the hybrid approach in seven organs. Differences were particularly large for the cochleas (0.78 versus 0.69), maculas (0.65 versus 0.00), and foveas (0.25 versus 0.00), showing that separate small-structure models were vulnerable to class imbalance.
- Performance remained weaker for several small or poorly visible structures, and no interobserver benchmark or dosimetric impact analysis was performed. The authors publicly released the model, processing code, and documentation, but clinical use still requires local commissioning, human review, and regulatory compliance.
CLINICAL TAKEAWAY
Combining contrast-enhanced T1-weighted MRI with planning CT appears more useful than adding T2 FLAIR for automated EPTN-based neuro-oncology contouring. The model is suitable for research and prospective validation, but it should not replace clinician review until externally validated and tested for dosimetric consequences.