Secondary use of radiological imaging data: Vanderbilt's ImageVU approach

David S. Smith, Karthik Ramadass, Laura M. Jones, Jennifer Morse, Daniel Fabbri, Joseph R. Coco, Shunxing Bao, Melissa A. Basford, Peter J. Embi, Reed A. Omary, John C. Gore, Jill M. Pulley, Bennett A. Landman

TL;DR

Vanderbilt University Medical Center (VUMC) has developed ImageVU, a comprehensive research imaging infrastructure that addresses the challenges of accessing and utilizing clinical imaging data for research. ImageVU enables secure, efficient, and regulatory-compliant access to imaging data, streamlining cohort discovery and supporting advanced research workflows.

What Problem Does This Paper Address?

The secondary use of clinical imaging data for research holds immense potential for advancing medical understanding, from disease characterization to precision medicine. However, several significant barriers hinder its widespread adoption. These include:

  • Data Silos and Fragmentation: Clinical imaging data is often stored across disparate and fragmented systems, making it difficult to access and aggregate at scale.
  • Technical Challenges: The DICOM format, large file sizes, and complex PACS (Picture Archiving and Communication System) environments present technical hurdles for researchers.
  • Operational Misalignment: Clinical PACS are optimized for diagnostic interpretation and real-time patient care, not for the high-throughput data access required by research. Attempting to serve both needs from the same system can lead to performance degradation.
  • Regulatory and Security Concerns: Handling protected health information (PHI) requires robust de-identification processes and adherence to strict privacy regulations, adding complexity to research data access.
  • Limited Metadata Utility: Clinical metadata, while present, is often inconsistently populated or not structured in a way that is easily searchable for research cohorts.

These challenges collectively limit the reuse of valuable imaging data, hindering translational research and innovation.

Methods

ImageVU tackles these challenges through a multi-component, purpose-built research imaging infrastructure. Key aspects of its architecture and methodology include:

  • Dedicated Research PACS: A separate research PACS (using DCM4CHEE) receives a real-time stream of clinical imaging data, creating a secure, research-focused environment without impacting clinical operations.
  • Ad Hoc Backfill Host: This component retrieves historical or missed imaging data that may not have been captured by the real-time pipeline, ensuring comprehensive retrospective cohort completion.
  • Cloud Storage System: ImageVU leverages Google Cloud Platform (GCP) for long-term, encrypted storage of imaging data, offering scalability and high-throughput delivery.
  • De-Identification System: A multi-stage de-identification pipeline is employed, utilizing the RSNA Clinical Trial Processor (CTP), custom filters, and manual review to ensure PHI is removed according to project-specific requirements. This allows for varying levels of de-identification, from full anonymization to retention of identifiers for identified studies.
  • Metadata Integration: Imaging metadata is extracted and integrated into VUMC's existing research data warehouses (RD for identified data and SD for de-identified data). This enables metadata-driven cohort discovery through tools like RD Discover.
  • Modular Design and Automation: The system is built with open-source tools and designed for scalability and regulatory compliance, minimizing disruption to clinical workflows.

Datasets

The paper describes the use of clinical imaging data acquired at Vanderbilt University Medical Center (VUMC). As of December 2024, ImageVU has processed:

  • 12.9 million MRI and CT series
  • 1.36 million studies
  • 453,403 unique patients

This data is used to support research projects, and the system is designed to handle large volumes of DICOM data. No new benchmarks or datasets were introduced by this paper; rather, it focuses on leveraging existing clinical data.

Key Insights and Ideas

  • The Power of a Dedicated Research Infrastructure: The core insight is that a purpose-built research imaging infrastructure, separate from clinical systems, is crucial for enabling scalable and compliant data reuse. This separation prevents performance degradation in clinical workflows.
  • Metadata is Key to Cohort Discovery: While imaging data itself is large and complex, well-structured and accessible metadata is the gateway to efficient cohort identification and selection. ImageVU's integration with VUMC's data warehouses is critical here.
  • Flexible De-Identification is Essential: Recognizing that research needs vary, ImageVU's multi-stage de-identification process allows for tailored privacy protection, accommodating different study requirements and regulatory approvals.
  • Leveraging Cloud for Scalability and Cost-Efficiency: Utilizing cloud storage provides the necessary scalability for large imaging datasets and can be cost-efficient, especially when combined with on-demand de-identification for research access.
  • Collaboration is Crucial: The success of ImageVU relies on close collaboration between research informatics teams, IT departments, and clinical stakeholders to navigate technical, operational, and regulatory challenges.

Key Results

ImageVU has demonstrated significant institutional impact since its inception:

  • Supported 75 Project Requests: Facilitating research across a wide range of studies.
  • Delivered Over 50 TB of Imaging Data: Providing substantial data volumes to researchers.
  • Engaged 55 Investigators: Broadening the reach of imaging research.
  • Contributed to 66 Published Research Papers: Highlighting the impact on scientific output.
  • Streamlined Cohort Discovery: The metadata-driven approach has significantly improved the efficiency of identifying research cohorts.
  • Ensured Regulatory Compliance: The robust de-identification and security measures help researchers meet compliance requirements.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Metadata Integration with SD Discover: ImageVU's imaging metadata is not yet fully integrated with the SD Discover platform, which currently limits direct cohort definition by researchers for de-identified datasets. Future work aims to address this.
  • Scalability Analysis of Open-Source PACS: While DCM4CHEE was chosen for its maturity, a full comparative scalability analysis of various open-source PACS solutions for research settings is an area for future work.
  • DICOMweb Standards: The current system utilizes DIMSE-based DICOM services. Future exploration of DICOMweb standards is planned to improve interoperability and enable modern web-based access, potentially facilitating cross-institutional collaborations.
  • Common Data Model Exploration: While ImageVU is currently focused on institutional-scale access, future work may explore alignment with common data models like OMOP MI-CDM to improve interoperability with federated analytics initiatives.
  • Storage and Performance Bottlenecks: The paper discusses challenges related to handling large volumes of small DICOM files and directory size limitations, which were addressed through caching and directory reorganization. Continued optimization in this area is ongoing.

Summary updated

BibTeX
@article{smith2025secondary,
  author = {David S. Smith and Karthik Ramadass and Laura M. Jones and Jennifer Morse and Daniel Fabbri and Joseph R. Coco and Shunxing Bao and Melissa A. Basford and Peter J. Embi and Reed A. Omary and John C. Gore and Jill M. Pulley and Bennett A. Landman},
  title = {{Secondary use of radiological imaging data: Vanderbilt's ImageVU approach}},
  journal = {J. Biomed. Informatics},
  year = {2025},
  doi = {10.1016/j.jbi.2025.104905},
}
Figure