Hubris and Hype | PACS Hack | Vector Flow Breakthrough

“I felt like the carpet was pulled out from under me, and I was left without the tools necessary to move forward.”

Canadian radiologist, Nancy Boniel, after being duped by malware-altered CT scans.

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.
  • Pocus Systems – A new Point of Care Ultrasound startup, combining a team of POCUS veterans with next-generation technology to disrupt the industry.

The Imaging Wire

Hubris and Hype
IEEE Spectrum published a pretty fierce review of “how IBM Watson overpromised and underdelivered on AI health care,” telling a story that combined impressive vision and solid marketing with a string of ambitious claims and big-dollar acquisitions, but yielded limited results so far. Here’s how IEEE Spectrum tells it:

The Watson Health story started in 2011 when Watson soundly beat a pair of human Jeopardy! champs and IBM almost immediately announced that the AI program will move on from quiz shows to creating an AI superdoctor. Watson Health initially focused on leveraging the NLP capabilities that made it so good at Jeopardy! but its efforts over the following years would reveal that healthcare wasn’t as ready to be mined as IBM would have liked (not journals, not EHRs, etc.). Despite these challenges, IBM continued making bullish AI statements and major acquisitions, while sometimes signing big AI deals, but not living up to its claims or delivering results from these deals. All of this resulted in a “hubris and hype” reputation problem for IBM that is still very present and not helped by articles like this.

Indeed, Watson Health is in a tough spot. Pieces like this make it around the internet faster than you can say “schadenfreude” and rarely give IBM credit for being among the first to try to bring AI to healthcare or acknowledge the fact that everyone else is still figuring out healthcare AI too. For example, the article describes IBM’s AI solutions as just “assistants that can perform certain routine tasks,” but doesn’t acknowledge that this is how most current healthcare AI solutions would be described.

Few would argue that IBM Watson Health could have used the time and money it put into AI more effectively and most would agree that IBM’s contribution to the AI hype bubble hasn’t been constructive, but IBM’s 5-8 year head start on most healthcare AI players and its massive business clout suggest that there could be “Watson Health comeback” stories in the future. Still, IBM may be wise to hold off on major claims or forecasts until that comeback is clearly underway.

A Vector Flow Ultrasound Breakthrough
A team of University of Arkansas in Fayetteville researchers were able to use vector flow ultrasound to create “detailed images of the internal structure and blood flow of babies’ hearts,” suggesting that this technique could improve cardiologists’ ability to diagnose congenital heart disease in infants. Using a BK5000 ultrasound with built-in vector flow imaging, the group successfully examined two pigs (one healthy, one with congenital heart disease), followed by a pair of three-month-old human babies (also one healthy, one with congenital heart disease), using the technology to image tissue and blood flow at a depth of 6.5 centimeters and identify abnormalities in the unhealthy child. This new technique could prove to be an important breakthrough for cardiac ultrasound (only pediatric for now), which has traditionally been valuable in measuring overall heart health, and may be able to provide much deeper insights through vector flow technology in the future.

Siemens and Intel’s CMRI AI Collaboration
Siemens Healthineers and Intel announced a collaboration, combining Intel’s CPUs and OpenVINO toolkit with Siemens’ MRI tech to develop an AI-based cardiac MRI segmentation and analysis model that applies “AI where the data is generated — right at the edge.” Siemens and Intel suggested that their combined solution would eliminate the labor and errors that are associated with traditional methods used to measure/segment CMR images without the “added cost or complexity of hardware accelerators” (take that, NVIDIA GPUs).

Intel isn’t the first chipmaker that comes to mind when most think of medical imaging AI, but it appears that the CPU giant is using collaborations and announcements like this to build out its AI capabilities and portfolio. This was also the case for Intel and Philips’ announcement last summer and it’s a valid strategy. There’s certainly an argument that AI has a home on the imaging device, or at least close to the device, and it’s going to take a solid combination of R&D and messaging to help prove that argument.

Researchers from Israel’s Ben-Gurion University got some of the radiology industry’s attention when they demonstrated an AI-based security system to protect against imaging device hacks, but just gained a much larger spotlight by developing malware capable of creating fake nodules/lesions (or removing real nodules) on CT or MRI scans. The malware would be installed on a PACS network (generally not encrypted), either physically or over the internet, and then access and alter scans as they pass through. That’s one way to create demand for medical imaging security.

The researchers tested 70 altered CT lung scans against a trio of skilled radiologists, who misdiagnosed 99% of scans with fake cancer nodules as cancerous and mistakenly found 94% of all scans with malware-removed cancer nodules to be healthy. Even after being informed of the malware, the radiologists misdiagnosed 60% of the nodule-added scans as cancerous and found 87% of the cancer-removed scans to be healthy in a different 20-scan dataset (10 altered). Meanwhile, the researchers were able to trick a popular lung-cancer screening software tool into misdiagnosing the scans with false tumors every time.

To defend against this type of hack, the researchers suggested that hospitals encrypt their PACS network, digitally sign all images, and set up radiology and doctor workstations to verify those signatures.

Envisioning ML’s Future
A new NEJM article by a trio of Google and Harvard AI experts revealed an optimistic forecast for machine learning in healthcare and gave some constructive steps to get us there. The authors shared a vision where ML helps to improve human and technological accuracy, enhances diagnosis and patient care with the help of massive medical outcomes databases, and even improves doctor-patient relationships through AI workflow efficiencies and expanding their access to clinical resources. That’s a pretty solid vision, and not out of the question, but they also see challenges that we’ll have to overcome to get there:

  • Data Access – The first step for machine learning is accessing large, diverse, and high-quality datasets, which has to overcome challenges in the areas of privacy and regulatory requirements, the technology barriers related to accessing clinical data from various institutions and data platforms, and the labor required to label datasets.
  • Bias – The team warned that AI developers and users should watch how bias affects training, particularly given vulnerable populations’ historically limited access to care, and therefore their underrepresentation in datasets.
  • Adoption & Regulation – The team emphasized that in order for ML to achieve mass adoption it would have to be regulated, covered with a clear legal framework, and supported with local infrastructure, similar to the network created for pharmaceutical distribution.
  • Discretion – Even if the above challenges are solved, the team warned that clinicians and patients should still use their best judgement on when and whether to follow ML-provided advice.

The Wire

  • Fujifilm announced the upcoming Japan launch of its new Synapse Sai Viewer, using AI to support diagnostic workflow on its Synapse 5 PACS system. The Synapse Sai Viewer launches with three homegrown diagnostic imaging workflows (CT image organ and bone extraction and labeling, temporal subtraction of CT bone images, a virtual CT image “thin slice” function) and Fujifilm plans to continue to add new AI PACS applications going forward. This seems like a pretty big milestone for Fujifilm, which has been among the most active AI investors, and will soon have a PACS AI app platform to show for it.

  • A report from the Health Care Cost Institute revealed that the share of outpatient healthcare service visits (vs. office service visits) increased from 11.1% in 2009 to 12.9% in 2017, correlating with a notable shift in Level 3 ultrasounds performed in outpatient settings (20.9% to 25.2% share). This shift comes with a significant healthcare cost impact, as the average price for outpatient Level 3 ultrasound visits increased by 14% to $650 through 2017 and average pricing for same scans grew by just 4% to only $241 in office settings. HCCI largely attributed these trends to the increased occurrence of hospitals buying private practices, turning office visits into outpatient visits and adding hospital “facility fees” to services that wouldn’t have require these fees before a practice was acquired.

  • Zebra Medical Vision landed its 3rd bone health patent, revealing that it has more bone health patents in the pipeline. The new patent focuses on the volumetric analysis of bone mineral density values from CT scans, while the two previous patents focus on estimating DEXA scores.

  • Konica Minolta’s Society of Breast Imaging booth brought the debut of a pair of new communication features that will be added to Exa Mammo PACS later this year. Exa Chat (allows users to securely communicate 1-to-1, with departments, or between hospitals and imaging centers) and Exa Peer Review (built-in workflow step to share and reference patients, studies, and approved reports, eliminating the need for 3rd party peer review systems) were previously introduced as additions to the Exa Enterprise Imaging platform at RSNA 2018.

The Resource Wire

This is sponsored content.

  • This Medmo video details how its healthcare marketplace platform and network of participating radiologists help underinsured patients pay as little as possible for their imaging procedures.

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!