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 nerve sheath of the vestibular nerve. While non-cancerous, its growth can lead to significant neurological complications. Clinical management currently relies on "maximal linear dimension" measurements from MRI to track tumor progression. However, volumetric measurement—calculating the total 3D volume—is recognized as a far more sensitive metric for detecting subtle growth and informing critical treatment decisions, such as Gamma Knife Stereotactic Radiosurgery (SRS).
Despite its benefits, manual 3D segmentation is labor-intensive and prone to inter-observer variability. While artificial intelligence (AI) offers a solution for automation, the development of robust models has been hindered by the lack of large, high-quality, open-access datasets. This paper addresses this gap by releasing a comprehensive clinical dataset and a high-performance baseline segmentation algorithm.
The Open Annotated Dataset
The released dataset consists of a curated collection from the Queen Square Radiosurgery Centre, featuring 242 consecutive patients undergoing Gamma Knife SRS.
The collection includes:
- 484 MR image sets: Comprising both contrast-enhanced T1-weighted (ceT1) and high-resolution T2-weighted (hrT2) scans.
- Expert Annotations: Manual segmentations established through consensus by a neurosurgeon, neuroradiologist, and physicist.
- Radiotherapy Metadata: Uniquely, the dataset includes DICOM-RT objects (RT Structure Set, RT Plan, and RT Dose), providing details on the actual radiation doses administered.
Methodology and Baseline Algorithm
The authors provide a baseline implementation of their 2.5D Convolutional Neural Network (CNN) framework, built using the MONAI (Medical Open Network for AI) open-source framework.
2.5D Convolutional Neural Network
Unlike standard 3D CNNs, which are computationally demanding and often struggle with anisotropic resolution (where slice thickness exceeds in-plane pixel size), the 2.5D approach effectively exploits varying resolutions. The network incorporates a computational attention module to help the model focus on the relatively small VS target within the larger brain volume.
High-Fidelity Discretization
A key technical contribution is the use of the Sunderland algorithm for converting manual contour lines into binary label maps (voxels). The authors identified that standard interpolation methods often fail to capture the "poles" (top or bottom slices) of a tumor, leading to missing data. Their chosen discretization strategy ensures higher fidelity to the original manual clinical segmentations.
Results and Performance
The automated segmentation algorithm demonstrated performance comparable to human experts:
- T1-weighted images: Achieved a mean Dice score of 94.5 ± 2.2%.
- T2-weighted images: Achieved a mean Dice score of 90.7 ± 3.6%.
- Human Comparison: These results align with the inter-observer variability between human experts (Mean Dice score of 93.82%).
The dataset spans a wide range of tumor volumes (0.04 cm³ to 10.78 cm³), proving the model's efficacy across diverse clinical presentations.
Discussion and Future Directions
While the dataset provides high-quality scans optimized for radiosurgery planning, further research is required to evaluate performance on general surveillance MRI scans, which may have lower resolution. The inclusion of radiation dose information also opens avenues for developing AI capable of automating radiotherapy planning. By providing this open-access resource, the authors aim to facilitate the development of more robust, multi-modal algorithmic models for precision neurosurgery.
Summary updated