Medical imaging system reference architecture - Healthcare Industry Lens

Medical imaging system reference architecture

This section describes key aspects of medical imaging systems, such as PACS and VNA solutions.

Diagram of a cloud-based medical imaging system architecture.

A cloud-based medical imaging system, such as a PACS or VNA, on AWS. High availability and low-latency study retrieval for medical imaging solutions.

  • The solution should be highly available and deployed across multiple AWS Availability Zones.

  • Medical imaging systems often consist of front-end viewers, application servers, databases, and storage for the imaging data. Where possible, each tier of the solution should be able to auto scale independently. Containerization or serverless can simplify operations. Auto Scaling based on load provides performance during peak demand and minimize costs during periods of low demand.

  • Images may be programmatically retrieved using the DICOM Message Service Element (DIMSE) or DICOMweb protocol. A Network Load Balancer may be used to route traffic on ports used by DIMSE.

  • End users likely demand low latency for retrieving and displaying medical images. Consequently, the data must be durably stored and highly available for immediate retrieval. Users may expect immediate retrieval of medical images that are several years old.

  • Recently ingested studies may be cached on SAN storage, EBS volumes, or high-performance file systems like Amazon FSx. Cost optimized solutions provision the minimum volume sizes needed to meet performance requirements, and maximize the use of cost-effective object storage like Amazon S3.

  • Medical image data tends to be accessed less frequently as it ages, so newly ingested data should land on Amazon S3 Standard, and then move to lower-cost tiers, such as Amazon S3 Glacier Instant Retrieval, as access frequency declines over time. Amazon S3 Intelligent-Tiering can automatically move data to the most cost-effective access tier based on access frequency.

  • Metadata for medical image objects and associated clinical data is commonly stored in a database. These databases may require high-performance storage for the requisite latency and IOPS. In-memory caches, like ElastiCache, may also be used to improve performance. Leverage fully managed database services to attain high availability with minimal operational complexity.

  • The data acquired by some medical imaging scanners — like MRI, CT, C-Arm — must be processed in image reconstruction to yield readable images. Image reconstruction can be thought of as a high performance computing (HPC) workload. Cloud based compute provides elasticity, reducing the time required to perform image reconstruction for emergency procedures.

  • Front-end viewers can leverage protocols like HTTP/2 to minimize image download times. Applications may also pre-fetch, cache, or prioritize transmitting the images that are likely to be opened by the end user.

  • On-premises caches can provide low-latency hot storage. AWS Local Zones and AWS Outposts may help meet hybrid architecture, latency, and data sovereignty concerns.

  • Redundant network connections between care settings and cloud services are recommended when a loss of connectivity can impact patient health. AWS Direct Connect should be used by customer sites with high study volume. Hybrid architectures may help meet stringent latency and business continuity requirements.

  • Data lakes can enable both operations and research and development. Datasets for the development of machine learning algorithms and AI features can be stored in data lakes. AWS AI services and Amazon SageMaker can help ISVs rapidly develop AI-based features drawing from a data lake. SageMaker Ground Truth can streamline the process of labeling data for model training.