The Cancer Imaging Archive

Vestibular-Schwannoma-MC-RC2 | Deep Learning Consensus-based Annotation of Vestibular Schwannoma from Magnetic Resonance Imaging: An Annotated Multi-Center Routine Clinical Dataset

DOI: 10.7937/bq0z-xa62 | Data Citation Required | Image Collection

Abstract

The “Deep Learning Consensus-based Annotation of Vestibular Schwannoma from Magnetic Resonance Imaging: An Annotated Multi-Center Routine Clinical Dataset” (Vestibular-Schwannoma-MC-RC-2) comprises 190 adult patients with unilateral vestibular schwannoma (VS), referred to King’s College Hospital, London, UK. Patients with neurofibromatosis type 2 (NF2) were excluded. Each patient has 1–8 longitudinal scans acquired from 2010 onwards, totaling 543 contrast-enhanced T1-weighted (T1CE) scans and 133 T2 scans across 621 time points. The dataset provides binary VS segmentations for 534 T1CE scans, along with demographic data (sex, ethnicity, age) and clinical decisions recorded at each time point. Segmentations were created using an iterative, consensus-based deep learning approach. This resource supports research on automated VS surveillance, tumour segmentation, longitudinal growth modeling, and clinical decision support.

Introduction

The Vestibular-Schwannoma-MC-RC-2 dataset is a comprehensive longitudinal collection of Magnetic Resonance Imaging (MRI) scans focused on VS. It includes detailed binary segmentations for each visible tumour on T1CE, facilitating the development and validation of segmentation and progression pattern analysis of VS.  

The dataset comprises MRI scans from 190 patients referred to King's College Hospital, London, UK, sourced from over 15 hospitals across Southeast England. All patients are over 18 years old and have been diagnosed with unilateral vestibular schwannoma. Patients with neurofibromatosis type 2 (NF2) have been excluded from this dataset. Patients with the other coexisting tumours were excluded.  

This dataset is crucial for enhancing reproducibility in research on VS. By providing comprehensive and routine clinical imaging data from multiple hospitals, it allows researchers to validate their findings across different clinical settings and imaging protocols. This is essential for confirming the robustness of automated VS tools. 

The dataset addresses significant gaps in existing VS datasets by including longitudinal data with up to eight time points per patient, compared to our previously published Vestibular-Schwannoma-MC-RC dataset with fewer time points. This longitudinal aspect enables the assessment of tumour progression and patterns, fulfilling a critical clinical need for continuous routine monitoring of vestibular schwannomas, despite the treatments patients undergo. Additionally, the clinical data provided in this dataset enable more comprehensive analyses by correlating imaging findings with patient demographics and clinical decisions. 

While the Vestibular-Schwannoma-MC-RC dataset primarily consists of T2-weighted scans, the Vestibular-Schwannoma-MC-RC-2 dataset focuses on T1 contrast-enhanced scans. This distinction allows researchers to explore different imaging modalities and their impact on tumour detection and progression. Additionally, the dataset includes scans from a different region of the UK compared to the Vestibular-Schwannoma-MC-RC dataset, which enhances the diversity and generalizability of the vestibular schwannoma data. Vestibular-Schwannoma-MC-RC2  dataset does not overlap with our previously published datasets.

Methods

The following subsections provide information about how the data were selected, acquired and prepared for publication.

Subject Inclusion and Exclusion Criteria

The dataset comprises longitudinal MRI scans from patients with unilateral sporadic VS, collected from over 15 medical sites across South East England, United Kingdom. A total of 226 patients were referred to the skull base clinic at King's College Hospital, London, where they underwent initial management between August 2008 and November 2012. Eligible participants were adult patients, aged 18 years or older, with a single unilateral VS. This included patients with prior surgical or radiation treatment but individuals with Neurofibromatosis type 2 (NF2) related schwannomatosis were excluded from the study.

All patients with MRI scans available for at least one time point were included in the study. Scans showing other tumours and those covering non-brain regions (e.g., neck) were excluded. Additionally, images with a slice thickness greater than 3.5 mm were excluded due to reduced sensitivity to small lesions and the impact of partial volume effects, which hinder accurate delineation and volumetric analysis of VS.

Data Acquisition 

The data were collected across multiple scans performed during routine clinical surveillance. To ensure reproducibility and transparency, MRI acquisition parameters are provided separately and grouped into the following categories:

• Scanner Details: Manufacturer, Manufacturer Model Name, Device Serial Number, Station Name, Magnetic Field Strength, Imaging Frequency
• Patient & Positioning: Body Part, Patient Position, Procedure Step Description
• Software & Protocol: Software Versions, MR Acquisition Type, Study Description, Series Description, Protocol Name, Scanning Sequence, Sequence Variant, Scan Options, Image Type, Deidentification Method
• Acquisition Timing & Numbering: Series Number, Acquisition Time, Acquisition Number
• Geometry & Image Orientation: Slice Thickness, Spacing Between Slices, Image Orientation Patient - DICOM, In Plane Phase Encoding Direction - DICOM
• Sequence & Scan Parameters: SAR, Echo Time, Repetition Time, Flip Angle, Coil String, Percent Phase FOV, Percent Sampling, Echo Train Length, Phase Encoding Steps, AcquisitionMatrixPE, Recon Matrix PE, Acquisition Duration, Pixel Bandwidth

Demographics and clinical information: 

The demographics and clinical data captures essential patient information and relevant standards for data collection. For each MRI time point, the following are recorded:

• Demographics: Ethnicity, sex, and age at the time of imaging.
• Treatment Decisions: The primary management strategies for vestibular schwannoma (VS) are included:
  • Routine surveillance
  • Stereotactic radiosurgery (SRS)
  • Surgery
    The majority of patients are managed with routine surveillance. For those undergoing active treatment (SRS or surgery), the treatment date is also provided.
• Prior Treatments: Information on any previous interventions is documented.
• Symptoms: The patient’s clinical symptoms at each time point are recorded.

This structured clinical information allows longitudinal tracking of patient outcomes and management strategies.

Data Analysis

The final curated dataset includes 190 patients, each with 1–8 longitudinal scans acquired from 2010 onwards, totaling 543 contrast-enhanced T1-weighted (T1CE) scans,  481 T1-weighted scans and 133 T2-weighted scans across 621 time points (mean 3.25 scans per patient, mean monitoring period 4.83 ± 3.08 years). All scan dates were uniformly shifted for privacy, with consistent offsets applied within each patient’s imaging series. Binary VS segmentations are provided for 534 T1CE scans; masks are not included for 9 post-operative scans with no visible residual tumour. Supporting data include demographics (sex, ethnicity, age) and clinical decisions documented at each time point.

After converting the original DICOM files to NIfTI format (.nii.gz), the following steps were applied to deface the patient scans. 

The defacing pipeline repository: https://github.com/cai4cai/defacing_pipeline.git

1. Reorientation of MRI Images
   The MRI scans are reoriented to match the MNI152 template orientation using the fslreorient2std tool. This step ensures consistency in spatial alignment across all images.
2. Skull Stripping
   The HD-BET (Hybrid Deep Learning Brain Extraction Tool) is employed for tight skull stripping, effectively extracting the brain from the surrounding non-brain tissue. This deep learning model automates brain extraction from multi-sequence MRI data.
3. Binary Mask Creation
   The skull-stripped image is then converted into a binary brain mask, delineating the brain tissue from the rest of the image.
4. Morphological Dilation of Brain Mask
   A 7 mm dilation is applied to the binary brain mask using morphological operations. This step creates a safety margin to ensure that critical brain features near the boundaries are preserved during defacing.
5. Loose Skull Stripping
   The dilated brain mask is applied to the original MRI image to generate a loosely skull-stripped version. This process removes non-brain tissue while retaining essential brain structures.
6. Affine Registration to Template
   The loosely skull-stripped image undergoes affine registration to a skull-stripped T1-weighted ICBM152 reference template. Mutual information serves as the similarity metric during this registration process.
7. Inverse Transformation of Defacing Mask
   The inverse of the affine transformation matrix is applied to a defacing mask, resampling it to align with the registered MRI image. This step prepares the defacing mask for subsequent application.
8. Application of Defacing Mask
   The defacing mask is applied to the registered MRI image, effectively removing facial features to protect patient privacy while preserving brain anatomy.
9. Manual Quality Assessment
   A manual review of the defaced images is conducted to assess the quality of the defacing process. This step ensures that the defacing is successful and that brain structures are intact.

Usage Notes

Data Organization and Naming Conventions

All imaging files are stored in NIfTI format (.nii.gz) and follow a consistent naming convention:

VS_MC_RC2_{studyID}_{mri_date}_{scan_type}.nii.gz

• studyID: An anonymized patient identifier (e.g., 001).
  
  
• mri_date: The uniformly shifted MRI acquisition date (YYYY-MM-DD format). Dates preserve the longitudinal spacing between scans.
  
  
• scan_type: Imaging modality, which may be T1, T1C (contrast-enhanced T1), T2, or T1C_seg (segmentation mask of the tumour on T1CE).

Examples:

• VS_MC_RC2_001_1991-04-28_T1C.nii.gz → T1CE scan
• VS_MC_RC2_001_1991-04-28_T1C_seg.nii.gz → Tumour segmentation corresponding to the above T1CE scan
• VS_MC_RC2_004_1991-07-21_T2.nii.gz → T2 scan
• VS_MC_RC2_001_1993-06-01_T1.nii.gz → T1 scan

Clinical Data

Demographic and clinical information for each patient is provided in VS_MC_RC2_demographics.xlsx. This spreadsheet links directly to imaging files via the combination of study_id and mri_date.
Variables include:

• Ethnicity, sex, age at MRI
• Symptoms (structured, with free-text expansion where relevant)
• Treatment plan (Surveillance, Stereotactic Radiosurgery, Surgery)
• Treatment date (for patients receiving active treatment)
• Previous treatment history
• No_residual (Post-surgery) flag indicating absence of visible tumour after surgery

Imaging Metadata

Detailed acquisition parameters are provided in VS_MC_RC2_Metadata.xlsx, which includes one row per imaging file. The column file_name matches directly to the NIfTI filename, allowing users to cross-reference and retrieve scanner, protocol, and acquisition details for each scan. Of the 1,157 scans included in this dataset, metadata is fully available for 1,022 scans (about 89%).

Practical Guidance

• Use the filename convention to link imaging scans with their corresponding segmentation masks.
• Match study_id and mri_date between imaging files and the clinical spreadsheet for longitudinal patient-level analysis.
• Use the metadata spreadsheet to explore acquisition variability across scanners and sites.
• Note that all dates are uniformly shifted for privacy but remain consistent across a patient’s longitudinal scans

 

Data Access

Version 1: Updated 2025/11/05

Citations & Data Usage Policy

Data Citation Required: Users must abide by the TCIA Data Usage Policy and Restrictions. Attribution must include the following citation, including the Digital Object Identifier: