Artificial intelligence could revolutionize medical imaging, but how can stakeholders translate the technology from research to clinical practice?

Countless studies have shown the many ways artificial intelligence could transform the medical imaging field.

From detecting lung cancer as well as human radiologists to predicting kidney functionfrom renal biopsies, machine learning and deep learning algorithms have demonstrated their potential to improve imaging diagnostics and patient care.

However, the promise of AI also comes with challenges. Unintended consequences can arise from using the technology, including patient safety issues, workflow disruptions, and inadvertent biases.

At a 2018 workshop, the National Institutes of Health (NIH), the Radiological Society of North America (RSNA), and The Academy for Radiology and Biomedical Imaging Research discussed the steps necessary to advance the use of AI in medical imaging. In a new report published in the Journal of the American College of Radiology, the group described the top priorities for bringing AI into everyday care.

“As with other new technologies that have been translated from initial research to widespread clinical practice, we need to recognize that there will be novel challenges for the clinical deployment of AI tools,” the report said.

“Understanding the nature of these new challenges, potential mitigation strategies, and a well-conceived research road map that ensure that advances in AI algorithm development are efficiently translated to clinical practice are of paramount importance.”

How can the healthcare industry accelerate the use of AI in medical imaging, and translate these algorithms from research to routine clinical care?

DEVELOP STANDARDIZED USE CASES

The group said that to date, use cases for AI in medical imaging have lacked standard inputs and outputs among comparable algorithms. Because algorithms may run on a local server or in the cloud, a standard way of accepting inputs and outputs for algorithms to process will be required.

“Without standardized inputs and outputs for AI use cases, it becomes challenging to develop standard data sets for training and testing, and in the end the resulting algorithms may show different results for the same finding,” the report said.

“Ideally, AI use cases should be developed using a format that converts human narrative descriptions of what the algorithm should do to machine readable language such as Extensible Markup Language or JavaScript Object Notation using clearly defined data elements.”

Structured use cases could create standards for validation before AI algorithms are ready for clinical use, the group said, and those in the medical imaging field could help develop these use cases.

“The medical imaging community, including medical specialties, academic institutions, and individual radiologists, can positively influence AI development by participation in the development of these structured use cases as well as creating the general standards and structure for AI specific use cases so that AI algorithms can be built with the same definitions and deployed in clinical practice in a consistent way,” the report said.

ESTABLISH DATA SHARING METHODS

In order to develop high-performing AI algorithms, models will need training on quality datasets that contain appropriate annotations or rich metadata. While there is a lot of innovation happening around this topic, the group stated that it is mostly concentrated at data-rich organizations, which can limit the widespread availability of information.

Privacy concerns can also limit the ability of institutions to make data publicly available, stalling the development of AI.

“Accelerating the release of publicly available data sets and AI techniques such as transfer learning that allow patient data to remain behind an institution’s firewall while exposing algorithm training to more diverse data may be able to help accelerate translation of AI into clinical practice,” the report said.

ASSESS CLINICAL AND INFRASTRUCTURE NEEDS

The report noted that there is a lack of user interfaces for bringing the results of AI algorithms into clinical workflows, which limits the deployment of AI models for widespread clinical use. IT developers will need to create an efficient user interface and user experience to integrate with existing clinical workflow tools to accelerate AI use.

Additionally, developers will need to establish vendor-neutral interoperability standards for communication between health IT systems.

“Understanding the infrastructure needs, including both qualitative and quantitative analyses for AI deployment in clinical practice—either local or cloud-based—will be critical in allowing use of thousands of AI algorithms in actual clinical practice,” the report said.

“The medical imaging community must be involved in assessing the clinical and infrastructure needs and work with existing standards bodies such as the National Science Foundation and the NIH Connected Health Initiative to find solutions that facilitate adoption of AI in clinical practice.”

ENSURE TECHNOLOGY IS SAFE AND ACCURATE

Stakeholders in the medical field should work with IT developers, government agencies, and the public to make sure AI algorithms are accurate, free of bias, and safe for patients, the group said. To do this, stakeholders will need to validate AI algorithms using datasets that contain demographic and technical diversity. Federal agencies such as the FDA have played a critical role in validating AI models to ensure patient safety.

“The FDA regulates a broad array of medical imaging devices as well as computer-aided diagnosis software and other algorithms that provide decision-making support to medical practitioners,” the report said.  

“The agency recognizes the rapid increase in digitization across the health care continuum and the importance of regulating computer software that is able to detect and classify disease processes and has been issuing regulatory guidance for software computer-aided detection and computer-aided diagnosis since 2012.”

Clinicians in the medical imaging field will be crucial to moving cross-industry partnerships forward.

“Creating models for validation and monitoring of AI algorithms and minimizing unintended bias will require collaborations between researchers, industry developers, and government agencies. The medical imaging community should play a leading role in facilitating these collaborations,” the report stated.

Although there are barriers to overcome, the use of AI in medical imaging holds a lot of promise. Going forward, industry stakeholders will need to work together to ensure the technology is safe, effective, and efficient.

“The future for AI applications for improved diagnosis in general and for image-based diagnosis is enormous. The opportunities and challenges summarized here can serve as a guidepost and road map for future development,” the report concluded.