Automatic segmentation of pediatric brain tumors in the posterior fossa and the surrounding anatomy: an efficient method for optimized preoperative planning and patient education.

OBJECTIVE

This study aimed to develop and evaluate an automated segmentation model for pediatric brain tumors in the posterior fossa region and the surrounding anatomy to facilitate 3D visualization of pediatric posterior fossa tumors for preoperative planning and patient education.

METHODS

Annotations were created for a total of 29 pediatric posterior fossa tumor patients, and an additional cohort of 32 patients was included for only tumor segmentation. Six nnU-nets were trained for automatic semantic segmentation of the tumor, cerebrum, cerebellum, brainstem, ventricles, and venous structures. Performance metrics were set at a minimum Dice similarity coefficient (DSC) of 0.85 and a maximum 95th percentile Hausdorff distance (HD95) of 5 mm to ensure clinically relevant 3D visualization.

RESULTS

The 6 networks created predictions for all patients in the test set. The median DSC and HD95 scores met the predefined performance criteria with median (IQR) DSC of 0.90 (0.093) for tumor segmentation, 0.97 (0.003) cerebrum, 0.97 (0.015) cerebellum, 0.91 (0.014) brainstem, 0.94 (0.038) ventricles, and 0.89 (0.077) venous structures. Overall high performance was found for automatic segmentation of the tumor and surrounding anatomy.

CONCLUSIONS

This study showed the successful development of an automatic segmentation algorithm tailored for pediatric posterior fossa tumors and relevant surrounding anatomy, enabling efficient 3D visualization for preoperative planning and patient education. This study provides a valuable basis for future research aimed at implementing advanced 3D visualization techniques for preoperative planning and patient education in pediatric neurosurgery.