Not all AI is created equal—experts caution against 'useless' applications
Experts are pushing back on the narrative that artificial intelligence applications are going to be the cure-all for the numerous pain points radiologists face.
In a recent research paper published by Current Problems in Diagnostic Radiology, experts contend that not all AI is created equal. They believe it is crucial to distinguish the “useful” applications from the “useless” ones.
“Useless AI comprises tools that replicate tasks radiologists already perform efficiently, offering little clinical benefit. In contrast, useful AI addresses tasks beyond human capability, such as radiomics, large-scale text processing, and high volume image analysis,” Teodoro Martín-Noguerol, MD, of the MRI unit at HT Medical—a leading provider of diagnostic imaging services in Spain—and colleagues noted. “By clarifying this distinction, we encourage radiologists to adopt AI judiciously, defend core clinical skills, and prepare for future technological advances.”
The authors used three measurable dimensions to define “useful” AI: 1) impossibility for humans—tasks that cannot be performed by radiologists due to inherent limitations, such as extracting radiomics features; 2) scalability—tasks that are possible for one case but become unmanageable when applied to thousands or millions of images or reports; and 3) demonstrable improvement in outcomes—through either diagnostic accuracy or workflow efficiency.
They acknowledge that there are numerous applications that shine at specific tasks and are beneficial in certain contexts. However, the authors pointed out that many app achieve what radiologists can already accomplish by themselves with ease, such as detecting a bone fracture on a single image, placing a lesion into a BI-RADS category, answering radiology board questions and explaining findings to patients. Unless an algorithm can outperform a human at each of these tasks on a wide scale, it is not helpful, the group suggested.
“Most of these tasks are part of our essence as radiologists and clinicians, including the skills acquired during our years of training. Do we really need AI for this? Do we really want to lose these features that make us unique? Where is the added value of using these types of AI tools?” they wrote, adding that, “over-reliance on automated interpretation, triage or report generation may gradually erode perceptual expertise and the nuanced pattern-recognition abilities developed through years of training.”
The group used radiomics as an example of the most useful form of AI. It is impossible for humans to extract metadata and feature information from radiology images outside of general changes in signal intensity or manual quantitative measurements from parametric maps. The amount of data AI can process is the key distinguishing factor that separates it from humans; for example, AI can sift through mammography exams with hundreds of images in seconds. This process is much more time-consuming for human radiologists.
“When we are talking about hundreds, thousands or millions of images or texts (including radiology reports and electronic health records), AI can and should have its place in our workflow, especially to alleviate our workload,” they charged.
The authors highlighted several additional examples of useful AI, detailing how it can benefit patients and improve workflow efficiency. However, organizations need to be leery with the applications they choose to invest in. As promising as some may seem, if they are handling tasks humans have already mastered, they are essentially “useless.”
“Useless AI could be closer to a technology that attempts to mimic what we already do well, rather than helping us do what we cannot,” the authors concluded. “If we are not careful, we run the risk of spending more time admiring these novelty-driven tools than advancing the real frontiers of radiology.”
Read more of what the authors had to say here.
In addition to her background in journalism, Hannah also has patient-facing experience in clinical settings, having spent more than 12 years working as a registered rad tech. She began covering the medical imaging industry for Innovate Healthcare in 2021.