AI Research Gets Clinical | cfMRI | Mammography Sweet Spot

“‘AI will assist us, not replace us’ Keep repeating that. Not realizing that it’s a Trojan horse. First get in the reading room. The technology will evolve with our help. Then eventually show us the door.”

A post on the Auntminnie.com forums providing a solid reminder of the resistance to AI that exists among radiologists. Or maybe it’s a solid reminder to radiologists about the threat of AI…


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


AI Research Gets Clinical

The radiology AI community woke up on Saturday to news of what some are calling “the first clinical trial in medical AI.” And get this, it didn’t involve radiology or come from a western-based research team. The study came from researchers in Chengdu, China who found that an AI solution from Shanghai Wision A.I. increases polyp and adenoma detection during colonoscopies.

Colonoscopy isn’t “our” type of imaging, but the fact that the trial was clinical, randomized, controlled, and assessed outcomes has folks in radiology AI excited. Here’s how it went down:

  • The team first held a performance study, achieving a solid 0.984 per image AUC
  • They then held a clinical study among 1,058 patients, randomized into standard colonoscopy (536) and AI-aided colonoscopy (522) groups
  • During clinical practice, the AI produced a “beep” if it identified a polyp during a colonoscopy and then allowed the endoscopist to view an AI-augmented video highlighting the polyp
  • The AI system significantly increased adenoma detection rates (ADRs, 29.1% vs. 20.3%), mean ADRs per patient (0.53 vs. 0.31), diminutive adenomas found (185 vs. 102), and hyperplastic polyps found (114 vs. 52)

It’s particularly notable that the researchers performed biopsies, as the act of removing polyps due to AI-based findings makes this a full-fledged AI clinical trial, and an early model for safety testing AI before use in actual clinical practice. The researchers are also planning to raise the study’s scientific higher ground a step further, performing a double blind study using a sham AI to correct any clinician bias that may have existed in this initial study.

Given the often emotional tone of many of today’s AI discussions, it’s studies like this that keep AI discussions focused on science and patient care (not threats and doubts), and that’s a good thing.


Cardiac Functional MRI

An international team of scientists announced the development of cardiac functional MRI (cfMRI), a new MRI method used to measure how the heart uses oxygen and heart muscle activity. The team is bullish about cfMRI, calling it “a new era . . . of doing cardiac stress testing to identify patients with ischemic heart disease,” that provides more detailed information much earlier without requiring imaging agents or physical stress testing (e.g. treadmills).

Next Up – After successful preclinical studies, the team is preparing to prove cfMRI’s clinical performance by measuring changes in blood flow to the heart while patients are connected to a breathing machine that changes carbon dioxide concentration.

Why They’re Bullish – Although previous research into oxygenation-sensitive MRI showed a high level of ‘noise’ with blurry images that they believe is from variations in the heart’s processing of oxygen, cfMRI features a new approach to average this variation and use it to study how the heart works.

Up Next – cfMRI may also be useful for other applications that require measuring heart blood flow (beyond coronary artery disease), such as measuring the effect of heart attacks or cancer treatments and future efforts to understand the role of oxygen in health and disease.


CAD’s Second Chance

A new editorial by Luke Oakden-Rayner in Radiology: Artificial Intelligence acknowledged radiologists’ rocky history with CAD, but assured that deep learning is bringing a different type of CAD to radiology, benefiting from far superior technology and decades of experience to better-inform “AI design, testing, validation, policy, and regulation.”

Because of this, Oakden-Rayner urged radiologists to embrace the new generation of CAD to help computers fulfill their clinical potential, allow radiologists to plan for the coming changes in their profession, and predict where this technology may succeed or fail so they can protect themselves and their patients. Given that many radiologists still view AI as more of a threat than a tool (see today’s quote of the issue), articles like this are necessary.


The Mammography Reading Sweet Spot

A study out of Norway found an accuracy sweet spot among radiologists who read between 4,000 and 10,000 mammograms annually and have read over 20,000 mammograms during their careers. The study looked at 2,373,433 mammography readings (6,634 with screening-detected breast cancer) from 121 radiologists, finding that:

  • Radiologists with annual volumes between 100 and 10,000 achieve similar specificity (87% to 90%) and SDC (4.9/1k to 4.7/1k)
  • However, sensitivity (81%) and SDC (3.9/1k) declined among radiologists with annual volumes over 18,000 (potentially due to fatigue)
  • False-positive rates achieved notable volume-based improvements, falling from 5.3% at 100 annual readings to 4% at 4,000 annual readings
  • Similarly, radiologists with 20,000 cumulative readings had far lower false-positive rates (3.6%) than those with 500 cumulative readings (6.7%), with improvements becoming more gradual after 20,000

Understanding that an annual volume of 5,000 is already viewed as the standard in many regions, this study still provides a solid guideline for annual and cumulative volumes.


The Wire

  • Speaking of mammography reading volumes, a UCLA and Cambridge team published a study that reveals their autonomous radiologist assistant (AURA) machine learning solution can reduce normal mammogram volume in radiologist workloads. Using data from 7,000 women who were recalled for assessment (5k for training, 2k testing), AURA achieved a 0.99 negative predictive value, identifying 34% of the negative mammograms among patients with a 15% cancer prevalence and identifying 91% of negative mammograms among patients with a 1% cancer prevalence.
  • A report from digital risk protection company, Digital Shadows, found that around 4.7 million medical-related files are openly available on the internet, including 4.4 million DICOM medical images (double the DICOM images the company found in 2018).
  • Siemens Healthineers released six security advisories, revealing that a range of its devices and solutions are vulnerable to a remotely exploitable Microsoft BlueKeep Wormable flaw (CVE-2019-0708) that could give attackers full control of a vulnerable device. Siemens recommended that users of a range of its products (including some imaging software and radiography devices) install a new Microsoft patch, perform certain security steps, or wait for forthcoming patches.
  • In an effort to compete for top tech talent and get closer with the local academic and pharmaceutical/ biotechnology community, Fujifilm Life Science announced the opening of its Strategic Business Office in Cambridge, Massachusetts. Although not directly connected to Fujifilm’s imaging business, this is a solid example of Fujifilm’s ongoing efforts to diversify in the medical field. Fujifilm Life Science certainly isn’t the first ship off to Cambridge, as Philips Healthcare announced a move to Cambridge last year and IBM’s Watson Health business moved to Cambridge in 2015.
  • Researchers from NewYork-Presbyterian and Duke found that delayed enhancement cardiac MRI (DE-CMR) should be the standard method for identifying infarct-related arteries (IRA) among patients with non–ST-segment–elevation myocardial infarction (NSTEMI – heart attacks due to narrow arteries), not coronary angiography. The study (n=114) found that coronary angiography could not identify the IRA in 37% of NSTEMI patients, but DE-CMR identified the IRA in 60% of those patients, while leading to new diagnosis in 46% of patients.
  • An expose from Kaiser Health News revealed that a growing number of hospitals are being accused of recruiting specialist physicians with “kickbacks” (unusually generous salaries and benefits) that challenge departmental profitability, but give the hospitals self-referral advantages. These hospitals are increasingly at risk of violating federal self-referral bans and anti-kickback laws intended to remove financial incentives that may influence physicians’ clinical decisions.

The Resource Wire

  • This Qure.ai blog post highlights how its qXR automated chest X-ray solution has helped scale tuberculosis screening to remote regions across the globe.
  • In this Carestream video, orthopaedic surgeon Dr. Bryan Den Hartog presents clinical images illustrating traditional CT vs. extremity CT imaging and discusses how the image resolution in the OnSight 3D Extremity System helps in his practice.
  • In this on-demand webinar, Nuance details how PowerScribe One frees the radiology team from distractions, interruptions, and delays that get in the way of staying focused, moving quickly, and working smarter.
  • How much does an MRI scan cost? According to Medmo, that depends. Scans made with the exact same device on the exact same body part could cost $225 at one facility and $2,500 at another. Medmo also provides some advice to make sure patients don’t pay too much for their scans, including using the Medmo Marketplace where the average MRI costs between $225 and $700.

Leave a Reply

Basic HTML is allowed. Your email address will not be published.

Subscribe to this comment feed via RSS

Get every issue of The Imaging Wire

Delivered right to your inbox. We will never sell or share your information.