por
John R. Fischer, Senior Reporter | October 16, 2017
A report by Sectra questions the value
of machine learning in radiology diagnostic
work
One of the most discussed topics among radiologists today is the appropriate integration and use of machine learning in their day-to-day tasks.
Now, a new report published by Sectra, asks radiologists upfront where the value of this technology lies in their diagnostic work, as well as where it is least valuable, and what concerns exist regarding its implementation.
“The most anticipated benefits are the use of machine learning for workflow support,” Daniel Forsberg, a senior research scientist at Sectra, told HCB News. “That is, how it can be utilized to automate various tasks that a radiologist commonly performs but which have little or no impact on their actual diagnostic work. For instance, automatic hanging of studies and series, user interfaces that adapt depending on context, and user support to protocol studies.”
Though radiologists are interested in the ways in which machine learning and artificial intelligence (AI) could possibly enhance workflow, concerns also exist, and make them skeptical of its use. Things such as fear of being displaced by AI, lack of reasoning as to how algorithms conclude their results, and the risk of algorithms producing false responses.
For the study, Sectra consulted radiologists in its key markets of Scandinavia, Benelux and the U.S. The physicians were asked to respond to a series of statements referencing the use of machine learning for tasks in a clinical example of a case of prostate cancer. Responses included completely agree, partly agree, partly disagree, completely disagree and neither agree nor disagree.
The statements addressed machine learning in the context of transferring data to online services for automated image processing, decision-making on what to read and when, the ability to make improvements in non-diagnostic workflow support, automatic annotation of organs such as the prostate, automatic characterization and scoring of lesions, decision support based on radiomics and population imaging, and periodic changes in system performance.
Automatic characterization and scoring of lesions gained much favor with 100 percent of participants either agreeing completely or partly. Machine learning was also received well in the use of automatic annotation of organs, decision support based on radiomics and population imaging, and applying improvements in the ability to provide non-diagnostic support.
Other statement responses varied, such as sending information to online services for machine learning-based automated image processing, which raised questions about the security of cloud-based services as well as its location and how data was encrypted.
Forsberg says the report is an indicator of the areas where radiologists would most like to see machine learning as well as an understanding of how invested they are in the implementation of such technology.
“Radiologists and other physicians have great expectations for AI and machine learning,” he said. “This tells me that both physicians and researchers will be willing to work together to implement, train and deploy AI systems that can provide real value in the clinical world.”
The discussion on the use of machine learning and artificial intelligence in radiology recently arose at the
SIIM-NYMIIS Regional Meeting in New York City, particularly in regard to whether they will replace radiologists over the next few years.