Guide

MedTech AI: An in-depth look at applications of AI in healthcare

The surge of AI innovations in MedTech underscore its potential in healthcare—while revealing the need to address technical, business, and regulatory hurdles

August 14, 2023

healthcare professional in scrubs holding a tablet

The healthcare artificial intelligence (AI) market was worth around $11 billion worldwide in 2021 and is forecasted to grow up to $188 billion by 2030—increasing at a compound annual growth rate of 37 percent. AI in medical imaging, a large segment of this market, was valued at $1.7 billion in 2022 and is projected to reach $20.9 billion by 2030. Put simply, the significant opportunity for AI in medical imaging is on par with that of AI in healthcare overall.

Medical imaging refers to the use of technology to view the human body in order to diagnose, monitor, or treat medical conditions; technology used to generate medical images includes magnetic resonance imaging (MRI) scans, ultrasounds, X-Rays, Computed Technology (CT), and other tools.

A wave of new AI/machine learning (ML)-enabled medical devices has come to market in recent years, with an 83% increase in approvals by the Food and Drug Administration (FDA) seen in 2021 compared to 2018.

To stay ahead of the curve and maximize market opportunity, MedTech companies must prioritize innovative data solutions and end-user optimization that supports the development of MedTech AI.

Here’s a look at the current state of medical imaging AI and where the industry is trending, as well as some regulatory risks and implementation/adoption challenges within the space that can impact the application of this technology.

Current use cases of medical imaging AI

More than 500 AI-enabled medical devices have received FDA approval and are available in the market today. These devices span a wide range of functions, but there are some use cases that are largely applicable to the medical imaging:

Image augmentation

AI models are being used to augment and deliver higher resolution images—and potentially even enhance captures such as 3D models for easier diagnosis.

Subtle Medical is an AI-powered software solution for positron emission tomography (PET) and MRI scans that increases the quality of the images produced and decreases the amount of time needed to complete the scan. The SubtlePET solution reports higher quality images conducted in 25% of the original scan duration, improving the patient experience and optimizing workflows to allow for more efficient use of scanners.
GE Healthcare’s AIR Recon DL is an AI algorithm that improves the quality of MRI images while reducing the overall scan time. The AI model removes noise and ringing from raw images and improves the signal-to-noise ratio, allowing for clearer images and up to a 50% reduction in scan times.

Diagnostic assistance

Medical diagnostics are being streamlined—and accuracy improved—through the integration of AI-enabled medical devices.

Medtronic’s GI Genius is an AI-assisted tool that physicians can use during colonoscopies to help identify colorectal polyps. The device works by identifying polyps in real-time during the procedure, allowing physicians to make faster and more educated decisions. The GI Genius device was recently used by Northwestern physicians who reported achieving a 13% increase in the detection and removal of colorectal polyps.
ScreenPoint Medical developed Transpara Findings, an AI model that can read Mammogram results and aid radiologists in the decision-making and detection process. Transpara Findings can quickly identify normal Mammogram results, allowing the radiologist to focus on the results that require greater attention. Transpara can also identify areas of concern on the image for the radiologist to review, bringing greater efficiency to the diagnosis process.

Process automation

AI can provide a more efficient approach to speech recognition software that minimizes redundancies for radiologists when dictating findings for reports. AI-enabled technology is providing efficiencies beyond speech recognition through automatic prioritization of cases based on need and radiologist availability, the optimization of workload imbalances across radiology departments, and real-time clinical intelligence and decision support.

Nuance developed PowerScribe One, a reporting tool for radiologists that is harnessing the capabilities of AI to reduce the time spent on reporting and allow radiologists to focus on patient care. PowerScribe One can help automate the reporting workflow, auto-populate reports to reduce errors and minimize redundancy, facilitate discrete data sharing across systems and platforms, and improve follow-up recommendation consistency with automated guidance support and quality checks.
3M’s M*Modal Fluency for Imaging is an AI-powered radiology reporting software that combines speech recognition technology with AI-driven real time clinical insights to reduce reporting time. M*Modal Fluency for Imaging improves the accuracy of reports through natural language understanding (NLU) technology, which allows the platform to better understand what the radiologist is dictating. The platform also increases reporting efficiency through the use of real-time clinical insights built within the reporting workflow to allow radiologists to identify and address documentation gaps.

Trends in medical imaging AI

AI continues to revolutionize the medical imaging landscape, and integration with Augmented Reality (AR) is emerging as a notable trend. AR, in conjunction with AI, facilitates the creation of immersive 3D models from medical images, thus enhancing a surgeon's understanding and manipulation of complex anatomical structures. A pioneering example of this is the Orsi Academy's application of NVIDIA Holoscan in the execution of in-human robot-assisted kidney surgery. This innovative approach is anticipated to become increasingly instrumental in advancing surgical precision and patient outcomes.

Another transformative trend in medical imaging is the convergence of AI, Precision Medicine, and Radiomics—the extraction of many quantitative descriptors from medical images. The integration of these three spheres offers significant potential to improve imaging interpretation, particularly in the context of complex pathologies such as brain tumors. AI algorithms can mimic neural networks, thereby facilitating tumor genotyping, the precise delineation of tumor volume, and the prediction of clinical outcomes.

An emerging field known as Radiogenomics seeks to link genomic features with imaging biomarkers of a given disease, thus personalizing patient treatment strategies. This integration also aids in early risk identification, such as pre-metastatic niche detection and is pivotal in the development of Theranostics, an approach that combines specific targeted therapy based on specific targeted diagnostic tests.

The role of AI in medical imaging is becoming increasingly significant in the context of emerging infectious diseases. With novel infectious diseases posing continual global health challenges, the need for rapid and accurate diagnostic tools is paramount.

AI can assist in automating the identification and interpretation of imaging biomarkers related to these diseases, thereby accelerating diagnosis, optimizing treatment, and potentially curbing disease spread. As research and development in this area continue, AI's role in infectious disease imaging is expected to grow, contributing to improved global health security.

Challenges

While applications of AI in medical imaging have the potential to greatly improve the diagnosis and treatment of diseases—especially in oncology—there are several factors posing challenges to the development and adoption of these technologies.

Limited access to data

Healthcare data tends to be siloed, making it difficult to aggregate and centralize in large quantities. This poses a significant problem to AI, as development teams need access to large datasets to create sufficiently sophisticated AI algorithms; AI can only detect or search for scenarios that have been introduced in the past. Without a large repository of data, the application will be inherently limited in its scope, which poses obvious problems to utility.

Data deidentification

Not all radiologic data uses universal format DICOM. Exceptions to DICOM are increasingly prevalent, which makes standardizing the deidentification of images more difficult to code for in AI algorithms.
Besides the coding challenge of deidentification, developers face additional complexities that must be addressed for certain use cases to be successful. Consider MRI images of the skull, for example: Common practices like applying defacing or skull-stripping algorithms are used to remove facial features, ensuring compliance with HIPAA regulations and preventing image-based identification. However, for development teams intending to employ this data for AI algorithm training, these practices pose significant hurdles. AI algorithms necessitate a substantial amount of data points, and when facial features are eliminated from the MRI images, the distinct differences between data points are reduced. As a result, training the AI algorithm becomes more generalized, potentially leading to a decrease in the model's accuracy.

Lack of clinician buy-in and education

While the number of innovative medical imaging AI solutions continues to grow, the adoption of these technologies has been slow among providers. One reason for this is a lack of knowledge of what these solutions can do, how they can integrate into workflows, and available evidence of clinical and operational efficacy.
Providers may be reluctant to adopt AI technologies until testing in real-world clinical settings across diverse populations has been completed. Prior to adoption, clinicians want to ensure that the tools produce positive results for all patients so that traditionally marginalized and underserved populations are not at a disadvantage.
Clinicians are more likely to adopt technologies that easily integrate into existing healthcare systems and clinical workflows so that there is no additional time or effort added to their day-to-day workload.

Computer power/cloud enablement of data limitation

AI models require significant computational power, often more than provider organizations are presently able to support. Companies like NVIDIA have developed chipsets that make progress toward solving this issue, but technology like this can be expensive and hard to come by due to the recent explosion of demand for AI tools.
The cost of migration to and storage in the cloud is very high and would be difficult for many healthcare companies to accommodate, especially when legacy data is involved.

Solutions: What’s needed to clear the main hurdles

Regulatory

Healthcare companies can prepare for the new regulatory standards of premarket assurance, pre-specifications, GMLP, and ongoing monitoring in the following ways:

First, companies should organize their IT operational practices around transparency, traceability, and accountability. They need to meticulously document their software versioning, data handling, and system updates. This documentation aids in demonstrating the system's performance and how changes impact that performance. Companies should also maintain a robust security posture to protect sensitive health data, aligned with healthcare-specific regulations such as HIPAA.
Second, healthcare companies should establish robust performance-monitoring systems designed to detect and remediate deviations from performance expectations. Some examples of monitoring options include Time-Series Analysis of model accuracy and precision, Real-time Anomaly Detection Systems, and setting system parameters to prevent worst-case scenarios—with regular reviews of performance by department leads and the ultimately the board of directors.
Third, companies must implement Good Machine Learning Practices (GMLP). The GMLP should encompass principles such as data management, model validation, interpretability, and robustness. For data management, it will be imperative that companies have data governance policies in place, including data quality checks, data anonymization for patient privacy, and proper data labeling practices. Model validation should be rigorous, with clear documentation of the validation process and results. Companies should also strive to make their models interpretable, meaning that they can explain how the model is making its predictions or decisions.
Finally, the model should be robust and reliable, performing consistently against key metric benchmarks and remaining aligned with the intended objectives under a variety of conditions.

User experience

Improving user experience (UX) and adoption for AI applications in medical imaging involves a multi-faceted approach that emphasizes usability, transparency, and meaningful interactions. Some key principles of effective UX include:

Simplicity and intuitiveness: Medical professionals often work under time pressure, so a streamlined and intuitive interface is crucial. The application should present the most important information in a clear, easy-to-understand manner. For example: highlighting areas of interest in the image, providing succinct explanations for AI findings, or enabling easy navigation through different scans or patient records.
User-centric design: The design process should involve users (e.g., radiologists, clinicians) from the outset to ensure that the application meets their needs and workflows. This can involve user interviews, surveys, or usability testing sessions. Feedback should be continually incorporated into the design and development process.
Training and support: Providing robust training and ongoing support can significantly improve user adoption. This can involve training sessions, detailed user manuals, and responsive technical support.
Performance and reliability: The application should load quickly, process images efficiently, and operate reliably to meet the high demands of the medical environment. Users should be able to trust that the application will work as expected every time.

By focusing on these aspects, AI solution developers can create applications that not only deliver powerful AI capabilities but are also easy to use and well-integrated into the healthcare environment, thereby improving user experience and adoption.

On prem vs cloud, buy vs build

Lastly, companies ought to perform cost/benefit analysis in order to address the conundrum of in-house vs. cloud for AI applications and choose the best tech stack for their use cases.

Doing so means performing a detailed total cost of ownership (TCO) analysis for in-house and cloud-based AI infrastructure. To account for costs, companies should not only measure the direct costs of hardware or cloud services but also indirect costs (such as maintenance, upgrades, internal computing costs, and security measures).

Regarding benefits, it’s necessary to first define strategic objectives and then consider how factors like scalability, flexibility, and potential improvements in AI performance and capabilities are optimized under the buy vs. build structure. In many cases, firms have outsourced infrastructure and platform layers cloud service providers due to superior back-end technology. Examples include healthcare AI innovators such the Mayo Clinic and Zebra Medical Vision have built their AI/ML platforms on Google’s Cloud tech stack.

Conclusion

The wave of innovations in MedTech AI, especially in medical imaging, is showing us the power of AI-enabled technologies in healthcare. At the same time, these advancements are highlighting technical, business, and regulatory challenges that must be considered and addressed for these technologies to have their full intended impact and continue to grow in sophistication and application over time.

To meet these challenges, MedTech companies and healthcare provider organizations alike need to focus on their alignment to evolving regulatory principles, end-user experience, and IT infrastructure. Ongoing attention to these three main areas is imperative for the continued advancement of the healthcare industry’s relationship with AI.

MedTech AI: An in-depth look at applications of AI in healthcare

Current use cases of medical imaging AI

Image augmentation

Diagnostic assistance

Process automation

Trends in medical imaging AI

Challenges

Limited access to data

Data deidentification

Lack of clinician buy-in and education

Computer power/cloud enablement of data limitation

Solutions: What’s needed to clear the main hurdles

Regulatory

User experience

On prem vs cloud, buy vs build

Conclusion

Let's Connect

Be Part of Our Team

Stay in the Know

Current use cases of medical imaging AI

Image augmentation

Diagnostic assistance

Process automation

Trends in medical imaging AI

Challenges

Limited access to data

Data deidentification

Lack of clinician buy-in and education

Computer power/cloud enablement of data limitation

Solutions: What’s needed to clear the main hurdles

Regulatory

User experience

On prem vs cloud, buy vs build

Conclusion