Anatomical position-guided auto-segmentation and dosimetric evaluation of vestibular schwannomas in gamma knife radiosurgery

Background: Vestibular schwannomas (VS), the third most common nonmalignant brain tumor, pose a significant challenge for geometric segmentation due to their irregular shape. The consistent anatomical origin of VS constitutes a domain-invariant prior capable of attenuating heterogeneity; however, contemporary studies have yet to exploit this characteristic. Purpose: This study investigated and externally validated a DLAS network for VS, leveraging the anatomical position guidance (APG) of VS origin to enhance geometric segmentation accuracy and efficiency. Method: A total of 345 contrast-enhanced T1-weighted MRI datasets from patients with VS were used: 242 public datasets for training, and 103 datasets (40 public and 63 institutional) for independent validation. The institutional dataset was additionally utilized for dosimetry evaluation. APG was constructed by aggregating the VS segmentations from the training dataset and integrated into three public networks as an input channel for region-of-interest cropping and post-processing. Segmentation performance was evaluated using the Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), Average Symmetric Surface Distance (ASSD), and Relative Volume Error (RVE). Dosimetric evaluation was conducted using plan parameters for Gamma Knife Stereotactic RadioSurgery (GK-SRS), including coverage, selectivity, gradient index, and D95%/99%. Results: The proposed APG significantly enhanced segmentation performance, although imaging characteristics were different in independent validation. Compared to the original DynUnet, the mean improvements of DSC, HD, ASSD, and RVE were 5.0%, 14.3%, 25.6%, and 41.1%, respectively. Additionally, the mean ± STD values of coverage, selectivity, gradient index, and D95%/D99% were 0.98 ± 0.05, 0.62 ± 0.19, 2.91 ± 0.19, and 13.5 ± 1.7/12.3 ± 2.0 Gy, respectively, comparable to clinical plans. Using a GPU, the average training time was reduced from 66.44 to 31.51 h, and inference time from 43.6 to 0.41 s. When using a CPU, the mean inference time was 14.5 s. Conclusion: This study showed that embedding a VS-specific anatomical information significantly enhances DLAS accuracy and efficiency, supporting routine GK-SRS planning even on CPU-only workstations. APG can be integrated into various deep-learning networks through channel-wise concatenation.