Introduction
While the core operations of clinical digital pathology are primarily concerned with the scanning and viewing of slides such that microscopes can be replaced with digital viewers, there are numerous additional software tools available to augment clinical operations within the digital pathology space. For example, algorithms can automate reporting or provide clinical decision support (CDS) to the pathologist or to the scan team for quality control purposes. The integration of these tools into the clinical workflow is complex and in many cases requires collaboration between multiple parties to build custom infrastructure in order to bring the tool online.

Methods
At Memorial Sloan Kettering Cancer Center, we have built custom middleware which provides a platform for the integration of viewing and AI tools to provide decision support, workflow enhancements, or other forms of automation. Specifically, we employ (1) a digital pathology database which gathers image metadata alongside orders data from the laboratory information system (LIS), (2) an API which provides an access point for downstream systems to request digital pathology metadata from the database, and (3) event-driven cloud-based architecture for subscribing to image scans and pushing images and clinical metadata downstream accordingly. This event-driven architecture allows for the filtering of image scans such that models or platforms of a specific scope may receive access to images and associated metadata in accordance with that scope, i.e. an immunohistochemistry model is able to receive immunohistochemistry scans while passing over hematoxylin and eosin scans. 

Results
Through this integration layer, we have achieved the integration of 4 distinct viewing and/or AI platforms which subscribe to and receive image files and clinical metadata in near-real-time. These platforms house 5 distinct AI models and generate inferences on the order of 1000 images per day.

Conclusion
By building custom middleware infrastructure, a platform can be created for near-real-time integrations of viewing and AI tools into the clinical ecosystem such that pathologists may interact with these tools for evaluation and use in clinical practice. This integration layer implements automated real-time decision-making around whether images are eligible for one or more available viewers or AI models, and for the automated routing of those images to the associated platform(s) to ensure the data is available on demand for clinical workflows.

Learning Objectives

1. Understand the complexities involved with integrating digital and computational pathology tools into clinical practice.          
2. Understand one approach which has been taken at a large academic medical center to address the complexities of digital tool integration.