Auto-Segmentation of Cranial Nerves III to VIII In MRI CISS Images
Abstract
Purpose
This study investigates a model for the auto-segmentation of cranial nerves in the context of radiosurgery treatment planning for acoustic neuroma and trigeminal neuralgia. Current treatment planning systems are unable to perform automatic segmentation for cranial nerves, and the delineation of these small nerves often results in prolonged manual effort. This research aims to improve the efficiency of contour delineation for cranial nerves, particularly CN III to VIII, in MR CISS images by employing nnU-Net.
Methods
Eight retrospective 1.5T CISS MRI scans were retrieved from patients who underwent stereotactic radiosurgery for acoustic neuroma or trigeminal neuralgia. Cranial nerve regions CN III, IV, V, VI, and VII–VIII region were manually traced and contoured by a radiation oncologist. The images were cropped into volumes of 100 x 100 x 100 pixels, centred on the pons. The segmentation model was trained with nnU-Net using 3D full-resolution configuration and 5-fold cross-validation over 100 epochs.
Results
The average Dice coefficients for the five structures in the training set reached 0.9 and above; however, performance for the testing set is lower, yielding coefficients of 0.47, 0.21, 0.64, 0.45, and 0.62 respectively. Corresponding mean surface distances measured are 2.12, 4.86, 0.40, 0.98, and 0.97 mm, indicating generally good positional accuracy. The substantial discrepancy in performance between training and testing datasets can be attributed to different contour precision and extension across patients depending on individual SRS treatment objectives, as well as anatomical abnormality resulting from patient conditions.
Conclusion
This study successfully demonstrates automated segmentation of cranial nerves in MR CISS images. Despite challenges due to limited sample size, small contour volume, and possible structural abnormality, the results demonstrated a good positional accuracy and visual agreement with manual contours, suggesting it to be a useful tool for enhancing treatment planning and evaluation efficiency after further development and validation.