By Dr. Ulrik Kristensen
It is already clear that AI will play a role in almost every aspect of medical imaging.
Operational analytics, data acquisition and dose monitoring, scheduling and workload optimization, automatic tissue segmentation and measurement tools, automatic image analysis and diagnostic reporting are all active targets for new AI tools. Machine learning image analysis algorithms are being developed for a diverse range of clinical applications, such as neurological disorder detection, lung nodule detection and breast cancer diagnosis. Until recently, these algorithms were developed mainly at university hospitals for internal research use only, as well as a handful of specialist image analysis companies. But with the AI startup community flourishing and the number of commercially available algorithms increasing, how these tools will be integrated in the radiologist workflow remains unclear. In this article, we review and discuss the different integration strategies for AI and propose how the algorithm deployment maturation will most likely materialize.
AI adoption and integration will depend heavily on the broader imaging IT structure and healthcare IT architecture. Today, radiology IT is still largely departmental. Hospital consolidation in the U.S. and some Western European countries is starting to drive enterprise imaging, and is creating shared resources between radiology departments within the same health system, or access to imaging data for departments outside radiology. On-premise data storage is, however, still the most commonly used architecture, though some hospitals in the U.S. and Western Europe are starting to adopt public cloud solutions as security concerns decrease and providers get more comfortable with public cloud storage, after some years of hybrid cloud for backup and disaster recovery.
The highly departmental structure of radiology and slow adoption of cloud has also had consequences for the development of AI, and initial development of radiology AI was limited to academic research projects within single hospitals. More recently, as the AI industry has proliferated outside of these early developers, it has remained limited by the confines of existing healthcare IT frameworks. This has led to a mixture of integration strategies today:
Although enterprise imaging is in demand, particularly in the U.S., PACS is still the most widely used viewing platform in medical imaging. Many AI vendors have attempted to make their algorithms easily accessible and integrated into the radiologist workflow by partnering with PACS vendors for integration into their software and clinical applications. The AI software would typically launch through a button in the PACS user interface; although integration could be even tighter, this is still a step forward compared to individually installed algorithms requiring separate login. A strong motivation from the AI vendors’ perspective is to tap into the PACS vendors’ distribution network while keeping the costs of doing business at a minimum.
AV or UV integration
For larger institutions, using advanced visualization (AV) software for diagnostic imaging, integration into the AV software is a different way of introducing AI into the radiologist workflow. Although it would make sense from a workflow perspective to have it integrated into the AV viewer used for diagnostic interpretation, this is not yet common. Integrated diagnostic viewing platforms with seamless PACS/AV user interface may be part of the future radiology workflow, but AV today is often used after initial screening on PACS, which means tools for initial triage and screening would currently make more sense to keep on PACS. Most AV vendors today therefore focus on AI tools for automatic segmentation and measurements being used further downstream after the initial screening has taken place.
Universal viewers (UV) have experienced significant uptake in recent years often as a central part of an enterprise imaging solution. With regulatory approval and a growing number of embedded AV tools, these viewers are increasingly being used as a diagnostic viewer, especially in small and midsize hospitals. As AI-based automated diagnosis for routine cases could benefit this segment, many enterprise imaging vendors are indeed looking at integrating AI into the user interface, either through direct integration for best-of-breed solutions or via an online AI marketplace for a selection of tools adaptable to individual clinical specialties.
As the number of available algorithms from AI vendors increases, the need for consolidated marketplaces for easier distribution, integration and access grows. EnvoyAI is one such marketplace, providing API-based integration of algorithms and partnering with AV, VNA and PACS vendors for integration. However, several of the major imaging vendors have also started creating their own marketplaces with integration into their user interface. In recent years, the differentiation of vendors into either platform vendors or clinical tool vendors has become more pronounced, and simultaneously, the number of partnerships between the two vendor groups increased to develop the capabilities of existing platforms. A marketplace for AI tools, therefore, fits well with the overall strategy of platform vendors, meaning they can focus less on individual tool development, and more on platform development and integration.
Short to midterm
The solutions being planned and prepared today point toward a mixture of integration strategies remaining in the short to mid-term. This is because each vendor is deciding their integration strategy for AI based on the strengths of their portfolio in each clinical application. For example, an enterprise imaging or PACS vendor with strong functionality, high case volume, and main customer base in cardiology, will typically choose tight integration of cardio AI tools into their platform, but use a marketplace for AI tools outside of cardiology where they have less expertise. Likewise, a PACS vendor with strong breast imaging skills will choose tight integration of breast imaging AI algorithms into their platform, but utilise a marketplace strategy for other applications. In the short- to mid-term, this will create a market where the most common clinical specialities, such as cardiology, breast and pulmonology, will see tight AI integration into diagnostic imaging platforms.
For AI vendors, such partnerships will allow access to a significant user base and power further innovation for these applications. However, AI vendors should not put their entire focus on tight integration with a single vendor, as enterprise imaging or PACS vendors outside their clinical specialty will also request their presence in other marketplaces, which could add an additional revenue stream and broader distribution channels.
Most AV vendors are currently focusing more on AI providing improved automatic segmentation and quantification to refine their solutions, building these in-house or through OEM partnerships and incorporating them completely into the existing software; but tight integration into the AV software hasn’t had much attention yet. A number of UV vendors are looking at integrating AI into their software, as in the longer term a demand for this could arise from the small to midsize provider segment.
Marketplaces will continue to develop in the short- to mid-term. While direct integration will be in focus for best-of-breed solutions within each clinical subject, marketplaces will increasingly be used for lower volume AI applications, and for getting the algorithms out to a broader audience. We will see the term marketplace being adopted by multiple vendors, even those with a very limited number of available algorithms. However, these marketplaces will evolve over time to include third party algorithms as well. Marketplaces will continue to be distributed through cloud for easy access and transition to OpEx business models such as volume-based pricing or SaaS.
In the longer term, best-of-breed integration will become tighter, with algorithms not only being in the same user interface, but in the same view, as we see with automatic segmentation and measurement tools in AV today. Accompanying this will be an industry consolidation where some of the AI tool companies within the main clinical subjects will be acquired by PACS or enterprise imaging vendors, often after long-standing partnerships.
Marketplaces will continue to be cloud-based, but with a bigger role for public cloud hosting companies. Traditional PACS-based analysis software will transition to thin-client with a proportion of the archiving on public cloud. As this will depend on regulations and local cloud readiness, it will happen at different rates country to country, with the U.S. leading the way. Looking even further ahead, this development in deployment models may lead to competition from vendors not traditionally associated with healthcare, as the cloud hosting vendors already hosting and computing the analysis may look into building their own AI algorithms. However, as AI becomes increasingly integrated into radiology solutions, these AI algorithms will add to the existing offerings but not cause dramatic changes in the market dynamics for PACS and enterprise imaging vendors.
With the prices of GPU processors decreasing, in the longer term we can also expect to see more machine learning at the edge built into advanced medical imaging modalities. Initially, this will be used for optimisation of image acquisition. As AI image analysis for diagnostics becomes more widespread, algorithms will be built into the acquisition hardware and used for real-time analysis directly during scanning, with the results visible and actively prioritising the radiologist’s worklist. There are significant benefits from doing the analysis directly during scanning. These include shorter scan times for triaged cases, prolonged higher resolution scanning if abnormalities are detected, and optimising acquisitions for incidental findings directly during scanning, thereby avoiding the need for follow-ups.
AI in reality
Recent industry news and press releases appear to suggest that AI in radiology is storming ahead, but the reality is that the siloed structure of radiology, poor cloud readiness and infrastructure, and challenging regulatory approval processes in most parts of the world are slowing down both development and adoption. Although the AI community has been very positive, AI has had to exist in a very complex and challenging sector. Healthcare is generally not a fast-moving market, and in many countries departmental PACS will continue to be the standard for years to come.
The diagnostic imaging platform will, therefore, be the main option for getting the solutions to many customers even in large mature markets such as Japan and South Korea, and marketplaces will be restricted due to little public cloud usage, and reduced to a small number of regulatory-approved algorithms in the short- to mid-term. The U.S. will continue to lead the way with marketplaces and a growing number of approved algorithms to choose from, and Europe will follow although regulations on cloud usage in healthcare will continue to be an inhibitor.
Side-by-side with this development and shift to cloud for image analysis, we will start to see edge computing with real-time analysis built into the advanced imaging modalities. However, if edge image analysis will primarily be used for image acquisition, helping to improve image quality, reduce dose and cut down repeat exams, or if it will slowly eat more and more into diagnostic image analysis, remains to be seen.
About the author: Dr. Ulrik Kristensen is a senior market analyst at Signify Research, a health tech, market-intelligence firm based in Cranfield, U.K. Signify Research is an independent supplier of market intelligence and consultancy to the global healthcare technology industry. Our major coverage areas are Healthcare IT, Medical Imaging and Digital Health. Our clients include technology vendors, healthcare providers and payers, management consultants and investors. Signify Research is headquartered in Cranfield, U.K.