Abstract Automatic segmentation of vestibular schwannomas (VS) from magnetic resonance imaging (MRI) could significantly improve clinical workflow and assist patient management. We have previously developed a novel artificial intelligence framework based on a 2.5D convolutional neural network achieving excellent results equivalent to those achieved by an independent human annotator. Here, we provide the first publicly-available annotated imaging dataset of VS by releasing the data and annotations used in our prior work. This collection contains a labelled dataset of 484 MR images collected on 242 consecutive patients with a VS undergoing Gamma Knife Stereotactic Radiosurgery at a single institution. Data includes all segmentations and contours used in treatment planning and details of the administered dose. Implementation of our automated segmentation algorithm uses MONAI, a freely-available open-source framework for deep learning in healthcare imaging. These data will facilitate the development and validation of automated segmentation frameworks for VS and may also be used to develop other multi-modal algorithmic models.
Introduction
Vestibular schwannoma (VS) is a benign tumor arising from the balance nerve. While non-cancerous, its growth can lead to significant neurological morbidity. Clinical management is increasingly shifting toward volumetric measurement, which offers higher sensitivity for detecting growth compared to traditional one-dimensional linear measurements. However, manual 3D segmentation is labor-intensive and subjective. To address the lack of standardized data for training automated models, this study introduces a high-quality, expert-annotated dataset and a robust baseline algorithm.
Dataset and Methodology
The released dataset comprises 484 MR image sets from 242 consecutive patients treated with Gamma Knife Stereotactic Radiosurgery (SRS). Each case includes a pair of images: contrast-enhanced T1-weighted (ceT1) and high-resolution T2-weighted (hrT2). Uniquely, the dataset includes Radiation Therapy (RT) objects, providing clinical context regarding treatment plans and dose distributions.
To protect patient privacy, images underwent a "de-facing" algorithm to obscure facial features while preserving internal brain anatomy. The ground truth segmentations were established through consensus by a neurosurgeon, neuroradiologist, and physicist. A critical technical insight was the use of the SlicerRT algorithm for discretization, which avoids the "missing slice" errors common in simpler 2D rasterization methods.
The baseline algorithm utilizes a 2.5D Convolutional Neural Network (CNN) implemented via the MONAI framework. This architecture processes stacks of 2D slices to capture 3D context efficiently. An integrated computational attention module allows the network to focus on the relatively small VS target within the larger cranial volume.
Key Results and Discussion
The baseline 2.5D CNN achieved a Mean Dice score of >93% for ceT1 images and ~90.7% for hrT2 images. This performance is comparable to inter-observer variability among human experts, suggesting the model has reached a human-equivalent ceiling for this task. The model also demonstrated robustness on post-surgical cases, which are typically more difficult due to altered anatomy.
The study reinforces the value of multi-modal synergy: ceT1 is superior for tumor visualization, while hrT2 provides better detail for surrounding organs at risk, such as the cochlea. Providing both modalities allows for more sophisticated, multi-input AI models.
Conclusion and Future Work
While the current model is highly accurate, its performance is currently tied to specific radiosurgery planning protocols. Future work must address domain adaptation to ensure generalizability across different hospitals and scanner manufacturers. Furthermore, the inclusion of radiation dose information in this dataset paves the way for the next frontier in medical AI: automated radiation treatment planning.
Summary updated