Adversarial AI | 4 Steps to AI Integration | Healthcare Disruption Coming

“If an AI tool can only address one (or even a few) types of abnormalities within a single patient, it will be of limited use for most imaging studies.”

University of Pennsylvania 4th year radiology resident and informatics fellow, Jeff Rudie, MD, PhD, on the shortcomings of today’s largely “narrow AI” solutions.

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

Adversarial AI Attacks
Medical AI found its way into the mainstream press again last week after a paper from a team of Harvard and MIT researchers outlined how machine learning algorithms can be tricked into performing “adversarial attacks.” These attacks start by making small changes to input data like “changing a few pixels on a lung scan,” that “could fool an A.I. system into seeing an illness that is not really there, or not seeing one that is.”

The researchers suggest that although there’s a potential that hackers will try to use this vulnerability to intentionally create misdiagnoses, it’s more likely that “doctors, hospitals and other organizations could manipulate the A.I. in billing or insurance software in an effort to maximize the money coming their way.” Kind of like AI-driven billing code manipulation.

This paper brings up an important point that AI diagnoses are only as secure as the images fed into them, but the likelihood that healthcare providers will hack their own AI tools to increase their billing is another question.

Spotlight on Contrast Agent Safety
The topic of CT contrast agent safety made its way to NPR last week, which reviewed contrast safety’s up-and-down history, and suggested that CT agents may be a lot safer than their reputation suggests. The contrast agent safety rollercoaster hit its peak (or nadir) in the early-to-mid 2000s when nearly all radiologists believed that contrasts posed a kidney damage risk (surveys: 100% of in 1999, 96% in 2006), contributing to the ongoing practice of withholding contrast with patients at risk of renal disease.

That’s until Columbia professor, Dr. Jeffrey Newhouse, discovered that none of the studies finding contrast agents to be dangerous actually used control groups (that’s Research 101 stuff), leading him to run a series of studies that he believes prove that although kidney damage is correlated with contrast-enhanced CT scans, they weren’t caused by it (that’s Freakonomics 101 stuff). Although many in healthcare aren’t in complete agreement with Newhouse that contrast agents are somewhere between “nearly harmless and totally harmless,” changing perceptions may be lowering the bar that defines when the diagnostic benefits of contrast enhancements outweigh their debatable risks.

The Four Barriers to AI Integration
Workflow integration, data variability, disease diversity, and clinical utility. These are the four barriers to AI integration in radiology practices, as outlined by a UPenn radiology research resident and informatics fellow in a recent ACR article. Here’s what that means:

  1. Workflow Integration – The first step in true AI adoption has to be seamless integration into a hospital system’s PACS and dictation software. Fortunately, this is understood by the PACS and dictation providers, but there’s still plenty of room to improve standardization and integration in this area.
  2. Data Variability – Unlike the clean and homogenous images used in research, real world data is less clean and more variable. Deep learning can overcome healthcare’s variability situation with “larger sample sizes and more diverse, unbiased training data.” However, getting that data is going to require continued and collaborative “standardization efforts for data anonymization, data sharing, and data annotation.”
  3. Disease Variability – Even with clean images and large data sets, most AI algorithms still have a “narrow” focus on a specific application. These “narrow AI” tools serve as an important first step in the evolution of AI, but they are “far different from the reality of a radiologist’s overall job” and ignore the “sheer diversity of human diseases.” For true integration, AI solutions must be integrated into more comprehensive solutions, covering every body part and every modality, and taking into account the fact that more than one disease or abnormality is often present in the same patient.
  4. Clinical Utility – “Even if we overcome the first three barriers . . . (AI) must fundamentally add value in order to be adopted into clinical practice and be worth paying for.” Specifically, AI tools must reduce time to report and act on critical findings, make radiologists more diagnostically accurate, and reduce radiologists’ time spent reading studies.

Philips and Carestream’s Customer Impact
OTech’s consistently valuable blog provided some early perspectives on Philips’ acquisition of Carestream’s Health IT division, revealing mainly positive reactions from Philips’ PACS clients and concerns among Carestream clients. The Philips clients’ positivity was largely related to Carestream’ enterprise archiving and storage capabilities, which they hope will replace Philips’ proprietary back-end, and improve their ability to integrate with VNAs (currently a difficult and unautomated process with Philips PACS). Carestream customers are concerned that the acquisition will bring unwelcome changes to their current processes and relationships (e.g. new contacts, changes to service and support) and that their current Carestream systems will become obsolete if it turns out that Philips’ main acquisition target was Carestream’s clients and channel, and not it’s technology (that is a possibility).

OTech has “some level of confidence” that this will work out for customers on both sides, but suggests that these customers “might want to ask for solid guarantees from (their) suppliers and keep all options open.” That’s good advice in any situation and especially situations like these.

A Change is Gonna Come
Change Healthcare’s 9th annual Industry Pulse survey revealed that many healthcare executives (n=185) are bracing for disruption brought on by new healthcare entrances (Amazon, Walmart, Google, Apple, etc.), processes, and technologies. A notable 32.2% of the executives believe this disruption will have the greatest impact on the industry’s standard business models, followed by changes to healthcare delivery (13.3%), the customer experience (11.1%), and the healthcare supply chain (8.9%), all of which are logical given the strengths of these “big tech” and “big retail” entrants. Recent M&A activity (e.g. CVS & Aetna, Amazon & PillPack) has 8.3% of respondents concerned about disruption from a new wave of “vertical one-stop-shop” healthcare companies, while an interestingly low 7.2% of the execs believe AI-driven disruption is on its way.

Future disruption is always a safe bet for any industry, although the level of attention that big-name disruptors like Amazon and Apple are getting is pretty interesting considering that their path into healthcare is still undefined. That said, the problems these potential healthcare disruptors are trying to solve (high costs, inefficiencies, lack of open standards) will exist regardless of how aggressively they end up pursuing healthcare and should be a target for any healthcare provider interested creating (rather than navigating) future disruption.

The Wire

  • Varex Imaging signed a deal to acquire Swedish linear array digital detector manufacturer, Direct Conversion, for €75 million ($84.7 million). That’s a hefty 4.6x multiplier compared to Direct Conversion’s $16 million in 2018 revenue. However, the company comes with over €40 million in multi-year agreements and Varex believes integrating Direct Conversion’s linear array digital detectors across its product line (including its CT detectors) will pay for the acquisition within three years, while expanding Varex’ addressable market by up to $500 million.

  • New research funded by Vidyo (grain of salt: they’re a telehealth video company) found that two thirds of healthcare IT executives (n=275) expect their telehealth budgets to grow over the next three years. Although remote monitoring of chronic patients led all current telehealth applications (64%), medical image capture/storage/sending was still near the top (54%).

  • The American Journal of Public Health published a study supporting the widespread belief that breast density notification laws (now in 37 states and federally required) lead to greater breast ultrasound utilization and cancer detection. However, these improvements are only true if the notification emphasizes the benefits of supplemental screening. The study looked at 1,441,544 screening mammograms performed in 2014 and 2015, finding that states that require notifications that emphasize the benefits of supplemental screening had 10.5 more ultrasounds per 1,000 mammograms and 0.37 more breast cancers detected per 1000 mammograms compared to states without density laws. Meanwhile, there was no significant difference between states with “generic” density notification laws (no emphasis on benefits) and the states without any notification laws.

  • Carestream announced that Shannon Medical Center of San Angelo, Texas (400 beds) installed two DRX-Evolution Plus systems and one DRX-Revolution Mobile X-ray system, while converting three mobile imaging systems to digital with Carestream’s DRX detectors. Shannon Medical Center purchased the systems from Dallas-based Carestream dealer, Southwest X-ray.

  • Researchers from Japan trained an AI algorithm that can identify benign or malignant masses in breast ultrasound scans more accurately than radiologists with two years of breast imaging experience and equal to radiologists with four years of experience. The algorithm was trained on 480 images of 97 benign masses and 467 images of 143 malignant masses and then validated against 120 images (48 benign and 72 malignant masses), outperforming the radiologists in sensitivity (95.8% vs. 58.3%-91.7%), specificity (87.5% vs. 60.4%-77.1%), accuracy (92.5% vs. 65.8%-79.2%), and diagnostic performance (91.3% vs. 72.8%-84.5%).

The Resource Wire

This is sponsored content.

  • In this Carestream video, an Orthopedic surgeon discusses how the OnSite Extremity CT has improved his business and the care he provides his patients.

  • 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.

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!