Storage, Compute, and Other Considerations
It is highly recommended to separate storage and compute components in an IT infrastructure setup. This separation allows you to manage and scale these components effectively and based on the growing needs of an organization. This section describes the most common storage types, their uses for a digital pathology lab, and some considerations that need to be taken into account for storage components of the IT infrastructure.
* Storage types
There are three main types of storage: (1) file storage, (2) block storage, and (3) object storage. We review these storage types below.
File storage on the network (aka Network-Attached Storage or NAS) provides a file-level data storage mechanism to enable access to a heterogeneous set of clients on the network. This type of storage is the main underlying IT component for storing and managing unarchived whole-slide images (WSI) captured from slide scanners.
Blocks storage on the network (aka Storage Area Network or SAN), on the other hand, provides a block-level data storage mechanism for consolidated data. This type of storage is not directly accessible via the local area network. SAN forms the underlying storage mechanism for applications that store block-level data, such as fully-managed large database services, etc. Block storage is inherently more expensive than file storage.
Lastly, object storage is a type of data storage architecture that manages data as objects. This type of storage does not provide file-level or block-level access to data. Object storage architectures can be used to store massive amounts of data and can be used to archive WSIs, report files in various formats, archived artificial intelligence (AI) models, and other unstructured data in a digital pathology lab. Object storage architectures are generally cheaper than file or block storage architectures.
* Storage capacity
Estimating the amount of storage capacity needed for a digital pathology lab depends on several factors and requires expertise and prior experience with digital pathology data types. However, using modern scalable IT infrastructure, organizations can start small and scale their storage capacity as they grow. Here, we have briefly mentioned some of the major factors that influence the storage capacity required for operations.
The vast amount of data volume stored on file storage servers is from unarchived WSIs that are being used by image viewers or IMS. Therefore, a rough estimate of the storage space requirement can be obtained by assuming a certain file size for each WSI and the number of unarchived WSIs at any given point of time (WSIs scanned with a 40x objective lens by a modern slide scanner, WSI file sizes rarely reach above 5GB). If WSIs in pathology lab are not going to be archived on other types of storage, sufficient file storage capacity needs to be provisioned to allow long-term storage of WSI over several years.
In a digital pathology lab with no AI services being used, having a block storage architecture is optional. Small databases used by IMS solutions can be hosted on file storage and a block storage is not required. However, digital pathology labs that are using (or intend to use) AI, it’s recommended to add block storage capacity to their IT infrastructure to allow high-throughput low-latency database transactions. AI-based medical devices can produce a massive number of database entries, and therefore, it is important to host the underlying database on a storage type that can provide sufficient performance to the lab users. As a rule of thumb, it is safe to assume that there will be up to 500MB of data produced by AI-based medical devices for each analyzed WSI.
To reduce storage costs, it is recommended to archive WSIs after a certain period of time or when the WSIs are not being used actively. These WSIs can be archived on object storage, which is generally cheaper than file storage. To estimate the required capacity of object storage, the number of WSIs that will be archived over several years and the object size (file size) of WSIs (usually up to 5GB for 40x scans) need to be taken into account.
* Data loss prevention
In addition to the above, digital pathology labs need to set up mechanisms to prevent data loss. There are several software and hardware configuration layers that would reduce the chance of data loss in an enterprise environment. Some of these measures/mechanisms are as follows:
- Strong authentication and authorization mechanisms
- Use of secure software solutions
- Appropriate data deletion policies on storage servers
- Backups, replication, and snapshots
- Encryption of storage and communications
There are many more considerations that need to be taken into account when IT infrastructures are being set up. Covering various aspects of data loss prevention is beyond the scope of this article. However, these configurations and policies need to be carefully discussed and implemented for digital pathology labs to avoid accidental data loss or data breaches.
In addition to storage servers, a modern IT infrastructure setup includes additional servers for compute requirements. In a digital pathology lab, for example, AI modules may be used to analyze WSIs and these analysis tasks require powerful compute capabilities built using Graphical Processing Units (GPU) and powerful processors (CPU) on servers equipped with large memory capacities. Some of the most important factors that determine the performance of a private cloud setup are addressed below. Since GPU configurations are extremely important for development and deployment of AI applications in digital pathology labs, this section primarily focuses on GPU requirements.
* Graphical Processing Units (GPU)
New variants of GPUs are being developed and released every few months, and therefore, mentioning specific GPU cards suitable for digital pathology labs would render this article obsolete in just a few months. However, there are factors that labs can consider when choosing the right GPU cards for their AI operations. Some of these factors include Single-precision and half-precision floating point performance (FP32 and FP16), number of processing cores, compatibility with AI development frameworks, GPU memory, and GPU cards connectivity. It is important to select the GPUs suitable for current and future AI deployments to avoid unnecessary hardware upgrades.
Moreover, some of the newer GPU cards released recently allow virtualized GPU profiles (vGPU) to enable better utilization of hardware resources when multiple processes or users are using the same GPU card.
Finally, opting for servers that support installation of eight or more GPU cards on a single server is highly recommended. High-density GPU servers reduce costs per GPU card, maintenance costs, and consequently, lead to lower TCO.
* Other hardware requirements
In addition to GPUs, compute servers need to be equipped with powerful CPUs and sufficient memory, as well as other hardware equipment. A digital pathology lab’s requirements for these hardware resources are quite similar to most other organizations that process data in a non-real time manner, and therefore, we will not cover them in this article.
Several other factors need to be taken into account when a new IT infrastructure is being designed. Detailed coverage of these factors are outside the scope of this article. However, we will briefly mention some of these considerations here.
One of the most important characteristics of a modern IT infrastructure is scalability, and therefore, it is of utmost importance to design the infrastructure to allow frictionless scalability for compute nodes, storage nodes, and other network equipment. “Scaling up” or “scaling out” are common terms that describe the two scaling approaches in IT infrastructures to address the growing needs of an enterprise. “Scaling up” refers to increasing the resources (storage, compute, etc) available in the existing network nodes, whereas “scaling out” means addition of extra nodes to the network in an incremental manner. Scaling out is generally considered to be the better approach in large enterprise-grade IT infrastructures and leads to significantly lower TCO over time.
Another consideration that can affect the TCO is the choice of operating systems. Compared to other operating systems, general-purpose Linux variants can provide a great amount of flexibility, security, and scalability in a network. Using the right tools to automate installation and maintenance of these operating systems is an important part of modern IT infrastructures. Moreover, it is possible to use container-specific Linux operating systems instead of their general-purpose counterparts to minimize the operations for maintenance and improve security and reliability.
Finally, organizations need to decide if they prefer to host the network equipment in server rooms located within the physical boundaries of their organization, or use “collocation” services instead. Building server rooms can be prohibitively expensive and lead to increased TCO. Appropriate equipment needs to be installed to maintain temperature and humidity within the acceptable ranges, and sound and heat insulation need to be addressed effectively. Several other equipment needs to be utilized in server rooms including air purifiers, physical access control to the room, CCTV, control and monitoring devices, power supplies and UPS, and fire suppression among others. Therefore, it might be a better option for digital pathology labs to consider collocation services, which allow them to rent server racks in a shared server room maintained by other companies in the area.
Building an IT infrastructure for a digital pathology lab requires planning, an in-depth understanding of the requirements, and technical expertise. Whether the requirements are going to be met using a public cloud setup or an on-premise infrastructure, it is important to design the architecture based on modern storage and compute models and be ready for future use cases such as the application of AI in pathology.