by John R. Fischer
, Senior Reporter | April 24, 2019
Medical imaging players now have a new ‘roadmap’ to go by for guidance in research and development of AI solutions to advance their field.
Developed from input collected at a workshop held by the National Institute of Health in August 2018, the report outlines key research themes and ways to advance foundational machine learning research for medical imaging.
"The potential value of these machine learning methods to medical imaging is a recent discovery. The workshop and the publications themselves are strong evidence that the key stakeholders are working together to set the agenda," lead author Dr. Curtis P. Langlotz, a professor of radiology and biomedical informatics at Stanford University and RSNA Board liaison for information technology and annual meeting, told HCB News. "I expect they will continue to collaborate as the agenda is carried out."
While expected to advance clinical imaging practice in a number of fields — image reconstruction, noise reduction, segmentation, computer-aided detection and classification, and radiogenomics, among others — research on the use of AI is still in its early stages.
The aim behind the workshop was to instill collaboration and collect feedback on how to enhance opportunities and the pace of research around medical imaging AI, in order to address gaps in knowledge and prioritize areas of study. The result was the report, which points to specific innovations for producing more publicly available, validated and reusable data sets to serve as evaluation criteria for new algorithms and techniques.
Among the points it highlights are:
• new image reconstruction methods that efficiently produce images suitable for human interpretation from source data
• automated image labeling and annotation methods, including information extraction from the imaging report, electronic phenotyping, and prospective structured image reporting
• new machine learning methods for clinical imaging data, such as tailored, pre-trained model architectures, and distributed machine learning methods
• machine learning methods that can explain the advice they provide to human users (a.k.a. explainable artificial intelligence)
• validated methods for image de-identification and data sharing to facilitate wide availability of clinical imaging data sets.
It also specifies that useful data sets should rely on methods for rapidly creating labeled or annotated imaging data. In addition, novel pre-trained model architectures specifically for clinical imaging should be constructed with methods for distributed training that reduce the need for institutes to exchange data with one another.
The authors specify that fulfilling these objectives requires greater collaboration among standards bodies, professional societies, governmental agencies, private industry and other medical imaging stakeholders.
"RSNA published the results of its AI Summit recently, which is well aligned with this roadmap. I have no doubt the other co-sponsors of the NIH Worshop, the ACR and the Academy, will be pulling in the same direction," said Langlotz. "I am hopeful that we are on the road to a well-funded AI research ecosystem, both in foundational and in translational research. Today’s AI research will transform tomorrow’s medical imaging practice."
The workshop was cosponsored by the National Institute of Health, the Radiological Society of North America (RSNA), the American College of Radiology (ACR), and the Academy for Radiology and Biomedical Imaging Research.
The findings were published in the journal, Radiology