2134 - Auto-Segmentation of Trigeminal Roots in MRI CISS Images for Radiotherapy Treatment Planning

Purpose/Objective(s):

This study aims to develop an auto-segmentation method for the trigeminal root specifically for radiotherapy treatment planning. Current treatment planning systems primarily utilizing CT images are unable to perform automatic trigeminal nerve segmentation. Manual segmentation is not only time-consuming but also labor-intensive. The proposed method targeted to enhance the efficiency of contour delineation and produce segmentations of the trigeminal root that are useful for dose estimation in radiotherapy planning.

Materials/Methods:

This study employs 1.5 T CISS MRI scans from 8 patients undergoing stereotactic radiosurgery for acoustic neuroma or trigeminal neuralgia. Trigeminal roots are manually contoured by radiation oncologists. To enhance the dataset, data augmentation is applied, resulting in a total of 200 volumes for training and validation derived from 5 subjects, while 3 subjects are reserved for testing. Leave-One-Subject-Out cross-validation is utilized to assess model performance.

A 3D U-Net architecture is implemented for the auto-segmentation task. To address the severe class imbalance, two 3D U-Nets with different loss functions are combined: the first 50 epochs employ weighted binary cross-entropy, followed by training with Tversky loss for the remaining 50 epochs. The resulting prediction map is converted to a binary map using a percentile threshold.

Results:

The average Dice coefficients for the training and testing sets are 0.69 and 0.45, respectively. The relatively low Dice indices are primarily attributed to the small volume of the trigeminal root and a general overestimation of the predicted volume. Nevertheless, the results exhibit good visual agreement with manual contours in terms of structure location and shape, with average Hausdorff distance being 0.52 mm and 1.27 mm respectively for training and testing set.

Conclusion:

This method successfully generates useful contours of the trigeminal root from CISS MRI for radiotherapy planning, demonstrating good visual agreement. Notably, the method does not require a large training dataset or complex model design, suggesting it as a straightforward tool to assist in contouring during radiotherapy planning.

Abstract 2134 - Table 1

	Dice coefficient	Average Hausdorff distance
Training set	0.69 (range: 0.67 - 0.73)	0.52 mm (range: 0.20 mm - 1.20 mm)
Testing set	0.45 (range: 0.41 - 0.48)	1.27 mm (range: 0.65 mm - 2.45 mm)