The rate of change in the field of oncology has continued to accelerate, with an information doubling expected to occur every two months – or expanding at the rate of 64x per year. The oncology field has long since crossed the line that individuals can keep pace with the rate of increasing knowledge. It shouldn’t be surprising that AI-based applications have simultaneously exploded within radiational oncology.

 

Through early 2023, the FDA had approved about 520 AI-based medical devices, with nearly 400 of them in imaging and radiational oncology. The rate of submittal and approval continues to expand. Over the past half dozen years radiomics and radiogenomics have evolved from infancy to sectors with significant technologies released into the commercial markets.

 

Motion management, various aspects of which have existed since the inception of the integration of imaging and radiation delivery, now requires integrated re-imaging of the cancer target. Workflows have been adjusted to account for tumor changes, anatomical changes in the patient due to food intake or weight loss.
Utilizing advanced radiomics can accelerate the reading of images by factors of more than 10,000, reducing the burden on radiologists, and shortening the time to incorporate the new imaging information into the treatment plan.

 

In 2020, Stanford released its first Quantitative Image Feature Pipeline (QIFP) platform, an open-source, AI-based quantitative image package supporting both planar and volumetric oncology images to create trackable biomarkers.

 

All these changes have had direct impacts on treatment planning (TP), with TP itself now heavily leveraging AI. In 2022 MIT followed Stanford’s QIFP with its own Holistic AI in Medicine (HAIM) framework, a generalized AI model supporting multimodality input to produce higher accuracy diagnosis and prognosis across a range of indications, useful in oncology TP.

 

Exacerbating the complexity of integrating all this new technology is the fact that the FDA has just released a suite of new regulations regarding the development, release, and field monitoring of AI, along with new regulations related to cyber security within and between all these different components. Even in an environment of more slowly moving technology, new regulatory guidance can be difficult to fully internalize and implement. The challenges within radiational oncology are significantly amplified due to the rapidity of new features being introduced into the market.

 

The reality of Addressing the New Radiation Oncology Market

 

The above is just a short discussion of some of the impact of AI on radiational oncology. It is inarguable that AI is expanding with startling speed across image acquisition, image analysis, diagnosis, treatment planning – already impacting every stage of the practice of oncological medicine. Many OEMs within radiation oncology have started development efforts to integrate AI capabilities within either their existing products or within next-generation products. For many manufacturers, the goal of integrating AI, however, is not a simple process. There are significant hurdles, technical, cyber security and regulatory each present real challenges to many organizations and taken together can represent risks to both the timeline of product introduction and the overall capability and competitiveness of the fielded products.

 

While OEMs in most cases recognize these risks, the rapid advances and competitive pressures have created a situation where they can’t afford to wait. Products within radiation therapy have extremely large, complex, and, often, aging codebases. Adding significant functionality by modifying workflows, acquisition pipelines, and data analytics modules within these large bases can be a very difficult endeavor. These efforts often reveal fundamental limitations in the original system architecture and design.

 

The question for OEMs can often be two-fold: is it even possible for my software system architecture to expand to include this new functionality and, if so, what is the most minimally invasive approach to re-architecting my product to include an AI engine that will likely need to be replaced every year moving forward.

 

Radiation delivery systems have exceedingly complex software control systems which are fully integrated with other parts of the imaging ecosystem, including planar and volumetric imaging systems, image analytics, treatment planning systems, not to mention the more mundane interfaces to PACS, EMR and HIS systems. The dizzing pace of AI capabilities’ evolution is significantly impacting both the operation of the delivery equipment as well the integrated workflows for assessing the results and planning follow-on treatment.

 

This represents one of the most difficult classes of technology to incorporate into medical equipment: fast-moving, new technology into slower- moving, complex older equipment within a new regulatory regime.

 

Specific Challenges to in Integrating Fast-Moving AI Engines – Either Existing or New Products

 

The fact that AI functionality is improving at an exponential rate is, in itself, an architectural challenge. The change in engine capability is measured in order of magnitude per year – far outside the range of anything that has been integrated into healthcare systems in the past. When the new technology is evolving at that rate, by the time they design-in a new AI engine and support algorithms, it will most likely already be obsolete. The first of the difficult decisions relates to evaluating not only the current state of different AI engines, but also predicting the rate of improvement of the various AI technologies. Betting on the wrong engine will have very negative impacts on competitive position in the very near future.

 

Related to this challenge is the need to architect the product such that it will be an easy task to excise and replace the AI engine while minimizing the stress on the overall system design, reducing the ripple effects into other areas of the system.

 

 

By properly architecting their products, OEM’s can effectively become “AI technology agnostic,” allowing the AI industry to continue to evolve at its own rate while periodically integrating “best in class” capabilities into their own product. To achieve this, OEMs must properly architect their software system.

 

The new AI capabilities must ultimately align to their system design fitting into the existing system within the context of the current architectural design. The AI capabilities should be integrated in an incremental manner; at first augmenting, then ultimately replacing current control and analytical functionality. By aligning the new functionality to the design concepts of the existing system it is usually possible to approach the redesign in a minimalist fashion. This process requires a detailed, objective assessment of the existing software from an architectural perspective. Architecture Tradeoff Analysis Methodology (ATAM) provides tools by which a system can be easily decomposed illuminating design tradeoffs either supportive or inhibitive to inclusion of new functionality. This assessment supports the process of creating a roadmap to approaches for enhancing the system design to support integration of an AI module (or various AI modules) into the system in a more ‘loosely coupled’ manner.

 

The Regulatory Impact

 

The technical effort of integrating AI is certainly material. The additional challenge related to new FDA guidance is also significant. The fact that radiation oncology requires a number of tightly integrated, but distributed systems exacerbate the cyber issue. The FDA has become very directive in the obligations put on OEMS – both technical and financial. The architectural analysis process discussed above will also be instrumental in more fully understanding potential cyber security issues within the product.

 

A bigger issue is the regulations associated with AI. The need for a Predetermined Change Control Plan (PCCP) requires a much more thorough approach to both the design of the system and the AI-based analytics. The PCCP provides the OEM the opportunity to present to the FDA at submission how they plan on addressing the inevitable evolution of the AI-based system. Once again, this may include the change of the engine itself, not just specific changes to the algorithm. A properly thought-out system, allowing for changes in both these areas may reduce downstream work by millions of dollars, all while accelerating the ability to evolve along with the technology.

 

Conclusion

 

This has been only a short discussion on both the potential and challenge of rapidly evolving AI engines. It is clear that the new technology provides both to oncology industry. If properly approached, the challenges can provide a significant market advantage for OEM’s, especially in the mid-to-long term.

 

Formal, objective assessment of both AI and product architectures will lead to better, more grounded plans, accelerating cost-effective integration, while moving product platforms into a position to better ride the coming rocket ship of AI improvements.