As the field of deep learning in medicine progresses from research to clinical deployment, practical considerations quickly become a primary concern for operational leadership. Hardware infrastructure, although a key enabler, presents unique challenges in the clinical arena, as it requires components such as GPU compute, high-speed networking, fast storage, and policies and procedures around usage.
By stepping through the typical project workflow at the MGH & BWH Center for Clinical Data Science (CCDS), this paper presents the following:
- The reasons for building such a system on-premises
- The challenges that must be confronted
- A case study in how such tooling is leveraged across the project lifecycle