Comparative Evaluation of Deep Learning Models for 3D Segmentation and Volumetry of Vestibular Schwannomas Using Large Heterogeneous Data Sets with External Validation.

BACKGROUND AND PURPOSE: 3D segmentation and volumetry of vestibular schwannomas (VSs) is a more accurate method to determine tumor growth on serial imaging, but manual annotation is time-consuming to implement in routine clinical practice. We evaluated and compared 5 deep learning-based segmentation models (nnUNet [Base, ResEncL], U-Mamba, UNETR, and MedSAM) for 3D VS segmentation and volumetry, and we examined the robustness to acquisition heterogeneity and generalization on an external cohort.

MATERIALS AND

METHODS

Our refined internal data set consisted of T1-contrast-enhanced images, including 2692 scans (n =383 patients) for training and 277 scans (n = 97 patients) for testing. Post-model training and validation, performance was evaluated on both internal and a publicly available external test set (n = 241) using the Dice similarity coefficient, maximum distance between the predicted and ground truth boundaries (Hausdorff distance), surface-to-surface (S2S) distance, and relative volume error (RVE). A subanalysis of the model performance was also performed to evaluate the impact of tumor volumes and data set heterogeneity.

RESULTS

The median Dice score on the external test set varied between 0.899 and 0.927 with U-Mamba achieving the highest performance, followed by nnUNet (Base and ResEncL). For these top 3 models, the median Hausdorff distance was 3.59 mm, while the 95th percentile Hausdorff distance was 1.6 mm. The S2S distance was <1 mm, and the median RVE (%) varied between 0.07 and 0.08. The median Dice scores were lower, 0.848-0.85, for smaller tumors (<200 mm3) and higher for tumors of >400 mm3 (median Dice score, 0.925-0.932).

CONCLUSIONS

Models based on convolutional neural networks, transformer networks, and foundational models show robust performance for VS segmentation. Given the consistently high performance and self-optimizing frameworks of convolutional neural network-based models (U-Mamba, nnUNet), these may be more suitable for clinical applications.