Trained to Train | AI’s Workforce Disruption

“As patients, if AI replaces radiologists, we will just have to accept it, just as we accept the use of an MRI scanner, a blood test, or a genomic test for a tumor.”

Duke University’s Maciej A. Mazurowski dismissing the argument that AI must gain patient approval in order to succeed.

Imaging Wire Sponsors

  • Carestream – Focused on delivering innovation that is life changing – for patients, customers, employees, communities and other stakeholders
  • Focused Ultrasound Foundation – Accelerating the development and adoption of focused ultrasound
  • Medmo – Helping underinsured Americans save on medical scans by connecting them to imaging providers with unfilled schedule time
  • Nuance – AI and cloud-powered technology solutions to help radiologists stay focused, move quickly, and work smarter
  • Pocus Systems – A new Point of Care Ultrasound startup, combining a team of POCUS veterans with next-generation genuine AI technology to disrupt the industry
  • Qure.ai – Making healthcare more accessible by applying deep learning to radiology imaging

The Imaging Wire

Trained to Train

Medical imaging machine learning models can be trained to train. That’s right. A new study in the Journal of Digital Imaging found that ML models can study radiologists’ attention levels and mammogram interpretation behavior and apply them to radiology training and teaching programs.

  • The Model – To train the teaching algorithm, the team had eight radiologists review 120 two-view digital mammography cases (w/ 59 cancers), capturing their search behavior using a head-mounted eye-tracking device and recording their decisions with separate software. The team also built a model to determine the type of missed cancer.
  • The Study – The researchers found that the model could identify signs of radiologists’ attentional level and decisions with 95% accuracy, noting that the model could improve transfer learning techniques by 10%. However, the separate model that was intended to determine the type of missed cancers was unsuccessful.

AI’s Workforce Disruption

A paper from Duke University radiology AI scientist, Maciej A. Mazurowski, argued that there is a “real possibility” that AI causes a significant disruption to the radiology workforce. Mazurowski gave balance his best shot, admitting that the future is unknown and acknowledging the many human components of radiology, but his main point was that there’s no evidence that the popular opinion that “AI will assist rather than replace radiologists” will prove to be true. Here’s the six arguments he refutes:

  • AI Will Never Match Radiologists’ Performance – It’s safe to say Mazurowski disagreed with this, pointing to computers’ mounting achievements and ongoing improvements, adding that “visual tasks are no longer ‘safe’ from the reign of AI.”
  • Radiologists do More than Interpret Images – Mazurowski estimates that radiologists only spend between 36% and “50% to 75%” of their time interpreting images, but countered with an estimate that radiologists’ procedures and consultation work “only” takes up 15% of their day.
  • Radiologists’ Efforts Would Shift to Patients and Other Physicians – Mazurowski bursts this bubble by asking why radiologists aren’t doing this already if it’s important enough to be a core part of their post-AI job.
  • The FDA Would Never Let Machines Do Radiologist Work – He argues that the FDA doesn’t have a rule for machine diagnosis and notes that the FDA is moving in the direction of more machine diagnosis.
  • Legal Liability Issues Would be Insurmountable – Mazurowski acknowledged this as a valid argument, but countered that lab tests are performed by automated systems without specialists reviewing each test.
  • Patients Would Never Put Complete Trust in Computer Algorithms – Noting that there haven’t been systematic studies on this topic, Mazurowski argued that healthcare consumers would generally prefer “whatever method of image interpretation is proven to be the most accurate” and technology choices aren’t generally made by untrained patients. “Nor should they be.”

Mazurowski’s good news for radiologists is that they play a crucial role as domain experts in the development of future algorithms, noting that “brilliant machine-learning scientists frequently devote their efforts to developing sophisticated solutions to unimportant problems.” However, if AI overtakes medical image interpretation it will still lead to a significant restructuring of the medical imaging professions, with roles shifting towards “overseeing the functioning of the machines, interfacing between the AI and the information recipients (patients, other physicians), and developing new uses of the AI-interpreted imaging modalities.”

This was a well written argument and Mazurowski provided some solid arguments that “AI assisting radiologists” is far from certain. However, one could similarly argue that AI’s imminent replacement of radiologists is also far from certain, or even that we’ll end up somewhere between these two futures.

The Wire

  • Researchers from Emory University identified “significant racial and ethnic differences” in ED medical imaging use. Of 225,037 adult patient ED visits from 2005 to 2014, white and Asian patients underwent diagnostic imaging in 51.3% and 50.8% of their visits, while black patients and patients of “all other races” received imaging in 43.6% and 46% of their visits (XR, CT, US, MRI). Black patients were also found to be less likely to receive advanced imaging (CT 0.80 OR, and MRI 0.74 OR) than white patients.
  • U.S. nuclear pharmacy association, UPPI, filed a lawsuit against Jubilant DraxImage, accusing the radiopharmaceutical company of a range of anticompetitive practices. The lawsuit alleges that Jubilant DraxImage acquired rival medications and withheld them from the market, increased the costs of two radiopharmaceuticals that it was the sole supplier of by between 500% and nearly 1,800%, and began forcing its customers to buy in bulk.
  • A Brigham and Women’s Hospital radiology report harmonization initiative was able to reduce radiology report template variations by a whopping 97%, while achieving 88% to 100% adherence to the new templates during the following nine months. Brigham and Women’s created an oversight committee, who developed template requirements, and then reviewed and harmonized all reports across the health system, keeping only 597 harmonized templates out of 19,687 before the initiative (237 of these harmonized templates were created during the initiative).
  • The ABR revealed that a relatively low 84% of radiologists passed the 2019 ABR Core Exam, representing the lowest pass rate for the key radiology test in the last seven years (previous range: 86.2% to 93.5%). Few in the radiology community gave ABR a passing grade for this year’s board results, calling the 15.9% failure rate suspiciously high, criticizing its elimination of the oral exam, and expressing general sympathy for the aspiring radiologists who just experienced this career setback.
  • A report from a Columbia University-based team confirmed that the U.S. was the world leader in radiology AI research publication volume between 2000 and 2018 with between 35% and 50% of all publications, although other countries are picking up the pace, particularly China (18%). Neuroradiology was the most-targeted AI subspecialty, followed by body and chest, and nuclear medicine.
  • Smithsonian Magazine recently explored the promise and potential pitfalls of healthcare AI, covering forecasts that healthcare AI “could democratize health care by boosting access for underserved communities” (both due to cost reductions and physician access), but warned that AI may also challenge patient privacy and be prone to bias. The article doesn’t exactly introduce new ideas to people inside of healthcare and AI, but it is an example of how Smithsonian Magazine’s educated and influential audience is being introduced to this subject.
  • Imperial College London researchers are developing a new imaging probe that could improve prostate detection during minimally-invasive keyhole surgery by generating ‘visual heat maps’ of prostate tumors. The new probe improves upon Lightpoint Medical’s robotic surgery probe, and if successful, would help clinicians remove all cancerous tissue.

The Resource Wire

– This is sponsored content

  • This Nuance case study details why the Hospital for Special Surgery chose to #ditchthedisk and how they benefited from moving to the Nuance PowerShare Network.
  • Did you know that imaging patients are most likely to no-show for their procedures on Mondays and Saturdays? By partnering with Medmo, imaging centers can keep their schedules full, despite the inevitable Monday no-shows.

You might also like

You might also like..

Select All

You're signed up!

It's great to have you as a reader. Check your inbox for a welcome email.

-- The Imaging Wire team

You're all set!