AI-Guided Image Acquisition in Ultrasound
By: Christina Nelson
The second day of the Food and Drug Administration’s (FDA) public workshop on the Evolving Role of Artificial Intelligence in Radiological Imaging focused on the benefits and risks of ultrasound as a candidate for AI-guided image acquisition. Compared to other imaging modalities, ultrasound presents the benefits of being low-cost and non-ionizing. In addition, advancements in miniaturization allow ultrasound to be used in a variety of non-traditional environments, including in ambulances and in the home.
During the public workshop, stakeholders discussed the following areas for applying AI in ultrasound:
1. Image acquisition
AI can be applied to guide the process in which the “six degrees of freedom” (6-DOF) motion of the ultrasound transducer would typically be guided by professionally trained users. Applying machine learning could reduce the learning curve that is often associated with accurately placing the probe. Probe placement is difficult for new users as image quality can be impacted by the way in which operators use the handheld probe.
Dr. Ha Hong, founding team member at Caption Health, introduced the first FDA-approved AI-based cardiac ultrasound guidance software during the workshop. Dr. Hong shared that it is possible to develop ultrasound navigation systems using AI that can be operated by users with varying levels of technical expertise. Caption Guidance does this by emulating the intuition of imaging experts by predicting the deviation of the current transducer positioning, providing the user instructions to minimize the deviation.
2. Image interpretation
Machine learning can also be applied in the analytic phases to help address some of the challenges that are typically associated with analyzing ultrasound images. Dr. Anthony Samir, Service Chief of Body Ultrasound Imaging at Massachusetts General Hospital, described that AI solutions can be developed in ultrasound to automatically place the region of interest, eliminate unnecessary imagery that is not valuable to the user, and provide AI-based computational diagnostics on the imagery. This can be useful for clinicians who are responsible for scanning and interpreting ultrasound images.
Regulating AI in Ultrasound
Concluding the opening session, the stakeholders participated in a panel discussion to examine the ways in which AI solutions can be effectively regulated by the FDA. The speakers identified the following opportunities for safeguarding safety and effectiveness:
Evaluation and testing
Monitoring algorithm performance pre-market and post-market will be fundamental to ensuring safety and effectiveness throughout the lifecycle of a device. Post-market surveillance of AI devices can help to recognize potential limitations that pre-market regulatory processes are unable to capture. In the case of no-human-in-the-loop AI, post-market surveillance will give insight into the risks that can occur when used in the clinical environment.
Establishing protocols that ensure algorithms are developed using appropriate, high quality imaging data can assist in developing quality AI devices. The concept of “garbage in, garbage out” was reiterated by several speakers, emphasizing the role that appropriate data plays in the overall success of AI devices. Imaging standards can be developed overtime that provide examples of high quality images with the correct resolution and parameters that can be used when training neural networks.
User education and training
Radiological devices are becoming increasingly sophisticated with the recent advancements in AI. As a result, user training that educates operators on the benefits and risks of ultrasound devices can help bring solutions into clinical use. In addition, it is important that manufacturers, users, and the FDA, support issues and accidents that may occur to prioritize patients’ safety.
Ensuring Patient Safety
The FDA Public Workshop started a thoughtful discussion on the many considerations that need to be made to bring safe and effective solutions to market. Evolving the existing regulatory toolkit will be driven by manufacturers, users, members of the FDA, and patients who are committed to accelerating progress by asking questions about how new solutions can be effectively evaluated and tested overtime. As Dr. Benny Lam, Principal Regulatory Affairs Specialist at Philips Ultrasound, concluded his talk on the benefits and risks of deep learning-enabled ultrasound:
“All of us are responsible for sharing the responsibility to ensure patients’ safety and well-being.”