Integrated digital pathology at scale: A solution for clinical diagnostics and cancer research at a large academic medical center

Peter J. Schüffler, Luke Geneslaw, Dig Vijay Kumar Yarlagadda, Matthew G. Hanna, Jennifer Samboy, Evangelos Stamelos, Chad M. Vanderbilt, John Philip, Marc-Henri Jean, Lorraine Corsale, Allyne Manzo, Neeraj H. G. Paramasivam, John S. Ziegler, Jianjiong Gao, Juan C. Perin, Young Suk Kim, Umeshkumar K. Bhanot, Michael H. A. Roehrl, Orly Ardon, Sarah Chiang, Dilip D. Giri, Carlie S. Sigel, Lee K. Tan, Melissa Murray, Christina Virgo, Christine England, Yukako Yagi, S. Joseph Sirintrapun, David S. Klimstra, Meera R. Hameed, Victor E. Reuter, Thomas J. FuchsJ. Am. Medical Informatics Assoc 2021 · DOI

TL;DR

This paper presents a comprehensive digital pathology (DP) ecosystem designed for a large academic medical center, integrating clinical diagnostics, research, and education. Their solution addresses the challenge of widespread DP adoption by providing a vendor-agnostic viewer and a robust data management system, enabling seamless workflows and facilitating large-scale cancer research.

What Problem Does This Paper Address?

Digital pathology (DP) promises to revolutionize how we diagnose and research diseases, offering benefits like faster turnaround times, remote access, and enhanced collaboration. However, its widespread adoption has been hindered by several challenges. These include the lack of integrated solutions that connect diagnostic, research, and educational use cases, the complexity of managing diverse scanner technologies, and the difficulty in compiling large-scale, de-identified datasets for research. Existing DP solutions often focus on single use cases and struggle with interoperability across different hospital systems and data formats. This paper tackles these issues by proposing a holistic DP ecosystem that addresses these integration and data management challenges at a major academic medical center.

Methods

The core of their solution is a vendor-agnostic, integrated digital slide viewer. This viewer is designed to be the central hub for accessing and interacting with digital pathology slides, regardless of the scanner model used. It's integrated into existing hospital systems, allowing pathologists to access slides from various sources (clinical, research, educational) within a single interface.

A key innovation is the Honest Broker for Bioinformatics Technology (HOBBIT). This system acts as a central data warehouse, systematically compiling and sharing large-scale DP research datasets. HOBBIT integrates data from digital slides, pathology reports, and clinical information, while crucially handling de-identification and redaction processes to protect patient privacy, adhering to HIPAA regulations. This includes automated date truncation, ID generation, and text redaction of pathology reports.

Datasets

The paper describes the use of a large-scale digital pathology dataset generated over three years at Memorial Sloan Kettering Cancer Center (MSKCC). This dataset includes:

  • 288,903 digital slides reviewed by 926 pathologists and researchers.
  • 3.6 million glass slides digitized in total, with a scanning rate of 140,000 slides per month.
  • Data includes anonymized images, redacted pathology reports, and clinical data from consenting patients.

The paper does not introduce new benchmark datasets but rather leverages the existing clinical and research data generated at MSKCC.

Key Insights and Ideas

  • Holistic Integration is Key: The success of DP adoption hinges on seamless integration across clinical, research, and educational workflows, not just isolated solutions. A single, vendor-agnostic viewer is crucial for this.
  • De-identification is Paramount for Research: The HOBBIT system's robust de-identification pipeline is essential for unlocking the research potential of clinical data while maintaining patient privacy. This is a significant hurdle that their solution addresses.
  • Vendor Agnosticism Reduces Complexity: Supporting multiple scanner vendors within a single viewer simplifies the user experience and avoids the need for multiple specialized viewers, a common pain point in DP adoption.
  • Clinical Validation is a Major Hurdle: Achieving provisional approval for primary diagnosis in New York State highlights the rigorous validation required for clinical DP systems.
  • Data Volume and Accessibility Drive AI: The ability to compile and access large, de-identified datasets is fundamental for developing and validating advanced computational pathology tools, including AI-driven diagnostic models.

Key Results

  • The system has been in operational use for over 3 years, with 926 pathologists and researchers accessing 288,903 digital slides.
  • 51% of digitally reviewed WSI were accessed within 1 month after scanning, indicating good adoption and accessibility.
  • The integration of the viewer into 4 hospital systems significantly increased DP adoption.
  • The system received New York State provisional approval in 2020 for primary diagnosis and remote sign-out, a critical step for clinical integration.
  • The HOBBIT system facilitates the translation of knowledge into effective health measures, including AI-driven detection models for various cancers.
  • The integration into research platforms like cBioPortal and PPBC has improved workflows and reduced the need for physical glass slides, with requisitions for glass slides decreasing by over 70% for PPBC.

Limitations and Future Work

  • Scanner Dependency for De-identification: While the viewer is vendor-agnostic, the de-identification of image data still requires vendor-specific processes, and the authors note that scanner vendors need to provide better de-identification software for their proprietary file formats.
  • Adoption Rates for Research Platforms: While clinical integration shows high usage, research integrations (cBioPortal, PPBC) have lower user numbers, suggesting ongoing efforts are needed to drive adoption in research settings.
  • Future Work: The authors plan to further enhance the system by adding features such as ordering additional stains from within the viewer, implementing clinical consultation queues, and enabling direct push of diagnostic reports and annotations back to the AP-LIS. They also emphasize the ongoing need for scanner vendors to provide better de-identification tools.

Summary updated

BibTeX
@article{schuffler2021integrated,
  author = {Peter J. Schüffler and Luke Geneslaw and Dig Vijay Kumar Yarlagadda and Matthew G. Hanna and Jennifer Samboy and Evangelos Stamelos and Chad M. Vanderbilt and John Philip and Marc-Henri Jean and Lorraine Corsale and Allyne Manzo and Neeraj H. G. Paramasivam and John S. Ziegler and Jianjiong Gao and Juan C. Perin and Young Suk Kim and Umeshkumar K. Bhanot and Michael H. A. Roehrl and Orly Ardon and Sarah Chiang and Dilip D. Giri and Carlie S. Sigel and Lee K. Tan and Melissa Murray and Christina Virgo and Christine England and Yukako Yagi and S. Joseph Sirintrapun and David S. Klimstra and Meera R. Hameed and Victor E. Reuter and Thomas J. Fuchs},
  title = {{Integrated digital pathology at scale: A solution for clinical diagnostics and cancer research at a large academic medical center}},
  journal = {J. Am. Medical Informatics Assoc.},
  year = {2021},
  doi = {10.1093/jamia/ocab085},
}
Figure