Automated Segmentation of Bilateral Vestibular Schwannoma in Neurofibromatosis 2 (NF2).
Authors
Affiliations (4)
Affiliations (4)
- Department of Otolaryngology-Head and Neck Surgery, University of California San Diego, San Diego, CA.
- Department of Radiology, Massachusetts General Hospital, Harvard University, Boston, MA.
- Department of Radiology, University of California, San Diego, San Diego, CA.
- Department of Otolaryngology-Head and Neck Surgery, Massachusetts Eye and Ear, Harvard Medical School, Boston, MA.
Abstract
Automated segmentation models for volumetric measurement of vestibular schwannoma (VS) have been developed for sporadic VS but not for bilateral VS. Automated segmentation would be especially valuable in this setting: patients with neurofibromatosis 2 (NF2) undergo numerous MRI scans, and automated analyses would aid in timely therapeutic decision-making. We developed a computer vision model for the volumetric measurement of bilateral VS. Eighty-seven individuals with VS (59 sporadic, 28 NF2) from our institution; 30 patients with sporadic VS from an open-source data set. A nnU-Net was trained on our institutional data augmented with the public data to develop an automated segmentation model for VS on T1-post contrast MRI. The model was tested on a holdout set of sporadic and bilateral VS scans. Dice score to compare pixel-wise agreement, qualitative review. Median tumor volumes were 0.345 cc for institutional sporadic VS, 2.05 cc and 0.501 cc for the larger and smaller tumors for institutional bilateral VS, and 1.36 cc for the public sporadic VS. There was a high incidence of comorbid intracranial pathology in the NF2 cases, including 39% with meningiomas and 32% with trigeminal schwannomas. The final model achieved a mean Dice score of 0.94 on the internal sporadic VS holdout set, 0.95 on public sporadic VS, and 0.87 on bilateral VS. On qualitative review of the NF2 cases, the model distinguished VS from adjacent non-VS lesions. The model also detected small tumors in cases with significant size asymmetry. The model struggled in cases with noncontiguous segments of tumor, often including one segment but not the other. This study is the first to report on the automated segmentation of bilateral VS in NF2. Further work is necessary to improve model performance, extend it to the postoperative setting, and apply it to other intracranial tumors.