Deep Learning Based Volumetric MRI Segmentation Algorithm for Vestibular Schwannoma Monitoring

Introduction: Vestibular schwannomas account for ~8% of primary intracranial lesions and are frequently monitored with long-term serial MRI. Accurate volumetric assessment of tumor size is crucial in both the preoperative and postoperative settings to guide clinical decisions regarding surgical timing, stereotactic radiosurgery, or continued surveillance. However, many vestibular schwannomas are small at presentation and subtle changes in total tumor volume can be overlooked when assessed by diameter-based measurements or visual inspection alone. In the postoperative setting, interpretation becomes more complex as deformation of residual or recurrent tumor tissue with the resection cavity may mask true tumor growth or mimic progression. For example, a small residual nodule may change in shape without a change in volume, leading to diagnostic uncertainty. Manual volumetric segmentation would be time-consuming, subject to inter-observer variability, and not feasible in clinical practice. Automated volumetric segmentation offers an opportunity to provide clinicians with standardized reproducible measurements of tumor volume and enable data-driven monitoring and treatment strategies.

Methods: We retrospectively collected 286 preoperative MRIs from 169 patients at our institution who underwent surgical resection for vestibular schwannoma. The dataset included 169 high-resolution T2-weighted (FIESTA/SPACE) scans and 117 T1-weighted contrast enhanced scans. Manual segmentations were provided by the study team for training and as ground truth for testing. Data were split into 80/20% training and testing cohorts. Our deep learning algorithm was a 3D SegResNet with single channel input, 32 convolutional filters, and 0.2 dropout rate. The model was trained for 3,000 epochs with dice as the primary performance metric.

Results: On the held-out 20% evaluation set, the model achieved a mean dice score of 0.851 ([Fig. 1]). Qualitatively, the model performed well on a diverse range of lesions including small intracanicular lesions, large koos grade IV lesions, and lesions having cystic or multinodular components. Of note, the same model takes either high resolution T2 (FIESTA/SPACE) or T1c for this performance, making it more generalizable across sequences.