This three-part blog series explores the following: 1) the current state of artificial intelligence (AI) and machine learning (ML) in focused ultrasound; 2) the ecosystem steps that are transforming other therapies and that may accelerate the development and adoption of clinical focused ultrasound over the next few years; and 3) the rapidly developing field of generative AI (gen AI), and the potential impacts of widespread gen AI adoption on focused ultrasound over the coming decade.
Focused Ultrasound + AI/ML Tomorrow: Maturing Preclinical and Clinical Capabilities
In the last quarter of 2023, we surveyed the community on the use of AI/ML in focused ultrasound and wrote about it in a blog post. Last month, in the first part of this blog series, we updated you on those usage patterns. Today, we cast our eyes toward the future of AI/ML in our field, and specifically, on AI trends among other therapies that can accelerate focused ultrasound adoption and improve patient outcomes over the next three to five years.
Developments in other interventions, such as radiotherapy, are instructive since AI/ML now assists with dose optimization, real-time imaging and adjustment, adaptive planning, and other areas. These uses reinforce our belief in AI’s potential to assist with focused ultrasound patient selection, treatment planning, and treatment monitoring as well. Each of these broad areas can be devolved into subcategories such as image analysis, adaptive treatment (i.e., “next best action”), quality assurance, clinical decision support, and workflow automation. But which technology trends could potentially mature AI/ML use within the focused ultrasound field? Several possibilities arise: focused ultrasound integration into AI healthcare platform plays, increased adoption of federated learning, and an embrace of synthetic data or data augmentation. Let’s explore each below.
AI Healthcare Platforms: Common Tools and Infrastructure to Accelerate Better Patient Care
“AI healthcare platform” refers to a comprehensive suite of technologies designed to develop, deploy, and manage AI healthcare applications. Such platforms integrate various data sources, such as electronic health records, genomic data, and – importantly for the focused ultrasound community’s applications – medical imaging and real-time patient monitoring data. AI platforms can also ensure data quality through cleaning and preprocessing.
The potential advantage of platform use in the focused ultrasound ecosystem is that it may reduce duplication of efforts by adapting solutions developed for other therapeutic interventions. AI healthcare platforms allow developers of therapeutic delivery systems and other experts to focus on their core strengths—whether their expertise is rooted in the science of focused ultrasound or in the intricacies of clinical point-of-care AI. In this respect, proprietary focused ultrasound solutions may be improved by integration with mature capabilities in image guidance, patient monitoring, target selection, and other related AI-enabled data-handling technologies.
The promise of AI healthcare platform opportunities can be seen in their recent proliferation, particularly in clinical workflow and point-of-care integration. Multiple solutions already exist, including AI-native solutions like NVIDIA Clara, multi-vendor integrated systems like RaySearch Laboratories’ RayStation, and specialized platforms like Mirada Medical’s DLCExpert. Taken in aggregate, they offer numerous benefits, including providing point-of-care (or preoperative) computational power and efficiency; comprehensive AI toolkits; scalability and integration; tooling for data management and security; collaborative ecosystems; and AI workflow automation. By harnessing the power of platforms, focused ultrasound therapies may mature quicker and at a lower cost than a “go it alone” approach would offer.
How can AI healthcare platforms be of value to the focused ultrasound ecosystem? The answer is multifaceted, but by leveraging mature AI/ML tools, the ecosystem may quickly converge on better patient selection, more accurate and effective treatments, and other gains. To expand on this, many ecosystem needs are shared across multiple vendors, and they parallel developments in radiotherapy interventions. Integration of vendor-specific software with AI healthcare platforms has the potential to enable or improve treatment planning decisions, determination of the next best action during procedures, and treatment monitoring:
- Treatment planning decisions. Integrating focused ultrasound treatment planning software with point-of-care platforms could allow physicians to develop personalized treatment plans based on historical treatment data and patient phenotype. The platform’s sophisticated use of predictive analytics could then enhance precision and treatment accuracy.
- Determination of “next, best action” during procedures. Whether in brain or body applications of focused ultrasound, determining the next best action during procedures is currently accomplished manually. Transitioning protocols to become “human in the loop” – wherein clinicians approve or adjust next actions as proposed by software – could significantly reduce treatment time and lower the risk of error. Automated adjustments to treatment parameters may also reduce physician workload while optimizing therapeutic delivery.
- Treatment monitoring. Detecting and imaging cavitation, and assessing the effects of focused ultrasound during treatment, could be assisted by digital twins and platform-attached simulations. By amalgamating advanced modeling, real-time data analysis, and predictive analytics, digital twins can simulate the acoustic fields generated by focused ultrasound waves, compare measured and simulated data, and predict when and where cavitation might occur. Such detailed understanding of the interaction between waves and tissue will ultimately drive better patient results.
Many AI-savvy companies have already developed analogous systems for other therapies. Collaborating with these companies could be more efficient than the focused ultrasound community creating similar AI/ML solutions, implying that close coordination between focused ultrasound manufacturers and platform developers would be advantageous. Beyond the business challenges, the first technical step is to map AI/ML applications relevant to the focused ultrasound ecosystem against the platform capabilities in existence today. Broadly, our field has diverse use cases that are not easily categorized into a single scenario or uniform data sources. For each application, generic processing pipeline diagrams would allow adaptation specifics to be determined.
Some companies possess therapeutic expertise while others have AI/ML expertise. By leveraging the benefits of partnering, while ensuring clear division of responsibilities to minimize regulatory and intellectual property issues, the field may advance more quickly.
Federated Learning: Collaborating While Protecting Patient Privacy
Federated learning (FL) is a collaborative approach that is used to train AI/ML models on decentralized data across multiple locations without sharing sensitive patient information. It protects patient privacy while advancing medical understanding.
FL is an important complement to clinical registries in understanding and predicting patient outcomes, and accordingly, it is improving the medical field’s ability to train AI/ML models to optimize patient outcomes. To understand both its allure and the problem that FL can mitigate, remember that AI/ML often drives inference, i.e., it may predict outcomes in many different forms. Can AI/ML help determine the patients to be treated or treatment particulars? Yes, of course. But a perennial challenge is that of data sharing to train models: patient data can be deidentified and aggregated to train models conventionally, but privacy concerns and the corresponding regulatory and legal risks are inhibitors. Finally, single-institution models often struggle to gather enough clinical data for scalable AI/ML training.
Federated learning enables the development of multi-institution AI/ML models while minimizing the privacy, legal, and administrative issues related to data sharing between these institutions. It allows participating parties, such as research hospitals, to collaboratively train robust AI/ML models while keeping data private. Scientific and administrative cooperation, and sequential (or parallel) model training results in versatile models that can predict outcomes across fluctuating conditions, including protocol variations, equipment differences, and changing patient populations.
Viewed holistically, FL can improve patient outcomes, protect privacy, and increase health equity:
- Improving Outcomes. FL develops consensus AI/ML models across institutions and protocols to predict optimal treatments, adverse events, and individualized treatment responses, directly benefiting patient health.
- Protecting Privacy. FL keeps patient data within the originating institution, which reduces privacy risks and encourages medical data donation.
- Increasing Health Equity. Because FL involves training across multiple institutions, with the widely available results, patients in rural areas can receive AI/ML-generated treatment plans that are comparable to those for patients in populous or affluent areas. Predictive models for uncommon conditions will now have sufficient training data, and subsequent interventions can improve outcomes for rare procedures.
Federated learning is transforming AI/ML in healthcare and shows promise as a valuable tool for improving patient outcomes. Alongside the creation and use of clinical registries, FL can also help us understand the nuances of complex interventions like focused ultrasound and should be part of our AI/ML toolbox.
Synthetic Data: AI-Generated Data to Safeguard Privacy and Accelerate Innovation
Synthetic data is another area where focused ultrasound can learn from other interventions. Synthetic data sets are generated through algorithms or computer simulations that replicate the statistical characteristics of actual data. In simple terms, AI-generated data mimics real-world data, and can be used to improve AI models, protect sensitive information, and mitigate bias.
The problem that synthetic data tackles is similar to that described above for FL: dealing with limited data sources where privacy concerns exist. Whereas FL’s advantages include the collaborative use of real-world data and corresponding data diversity, synthetic data solves these problems with an emphasis on increased data availability and simulation of rare events.
The challenges in the field of focused ultrasound that could be addressed in part by synthetic data adoption include the lack of real-world clinical data, the need for patient consent to consolidate real data, and the high costs of data collection and protection.
- Lack of real-world clinical focused ultrasound data. Synthetic data creation and use can help overcome a lack of real-world clinical focused ultrasound data.
- Patient consent and privacy. Clinical registries are an important tool for determining standards of care, but they require patient consent for data consolidation, and data deidentification may provide varying degrees of privacy. Synthetic data can be used to train models without the need for “opt-in” agreements or data movement.
- High costs and time required to collect real-world data. Synthetic data creation and validation (e.g., for accuracy, consistency, and reliability) requires an upfront expense but is usually cheaper to create and use than its real-world equivalent.
From a patient perspective, well-trained AI/ML models offer the potential for improved clinical experiences. Synthetic data could speed the research and development process, allowing for improved models and ultimately save both time and lives.
As a side note for technologists, many of these principles also apply to data augmentation. Whereas synthetic data comprises entirely new data created from scratch, data augmentation entails expanding upon existing data to enhance the size and variability of the data set. We have simplified the terminology in this section for brevity, but both needs exist.
Conclusions
Today the biggest inhibitors to growth in the focused ultrasound field are not technological but are perhaps due to a lack of awareness and ecosystem investments. As AI/ML become increasingly commonplace, however, its sophisticated use trends toward becoming a minimum requirement for therapeutic solutions. Therefore, we must collectively master and deploy AI/ML solutions in the coming years.
Fortunately, small focused ultrasound projects have already begun advancing these objectives. Importantly, we must always ask the question, “How does this action benefit patients?” Rather than be swayed by technology for technology’s sake, we should ensure that any project has the potential to improve care paths, save lives, or reduce costs.
In the coming years, AI/ML technologies will bifurcate. On one hand, they will drive exciting new highly visible possibilities for patient care. On the other, they will gradually disappear from our consciousness, fading into the background as so many other technologies have done after widespread adoption. Our job as clinicians, engineers, scientists, and business leaders is to decide where to apply energy and resources to drive the most value for the patients.
Rick Hamilton is the Foundation’s managing director and chief technology officer. He is recognized as one of the most prolific inventors in world history, with over 1,060 issued US patents.
Read the Blog Series
Part 1: Focused Ultrasound + AI/ML Today: The Intersection of Two Rapidly Evolving Fields
Part 2: Focused Ultrasound + AI/ML Tomorrow: Maturing Preclinical and Clinical Capabilities
Part 3: Generative AI and its Future Impact on Focused Ultrasound