Uncategorized

Gotcha Research | Data-Centric AI | Akumin Expands


” . . . there will be a lot of twists and turns on the way, but we will get there.”

Andrew Ng on whether he still believes his 2017 prediction that AI will become “the new electricity.”


Imaging Wire Sponsors

Arterys | Bayer Radiology | Canon Medical Systems | GE Healthcare
Healthcare Administrative Partners | Hitachi Healthcare Americas
Novarad | Nuance | Riverain Technologies | Siemens Healthineers
United Imaging | Zebra Medical Vision



The Imaging Wire


UW & Aidoc’s Performance Problem

A University of Wisconsin study called out Aidoc’s C-Spine cervical spine fracture detection tool’s “poor diagnostic accuracy,” using these results to highlight concerns about AI’s overall “generalizability, utility, and rapid pace of deployment.”

The Study – The researchers used Aidoc C-Spine to retrospectively identify fractures in 1,904 non-contrast cervical spine CTs (both acute & non-acute), comparing the results to a UW neuroradiologist’s findings.

The Results – Although Aidoc C-Spine and the neuroradiologist had the same interpretations in 91.5% of cases, the algorithm correctly identified just 67 of 122 fractures (54.9%) and flagged 106 false-positives.

To be Fair to Aidoc – Most of the cervical spine fractures in the study were non-acute, and the researchers found that Aidoc C-Spine was “significantly more successful at identifying acute fractures.” That’s good news for Aidoc, considering that C-Spine is intended to triage emergency patients for acute cervical spine fractures, not non-acute fractures (chronic or pre-existing).

UW & Aidoc History – If this study seems familiar, it’s because UW published a similar study three months ago questioning the accuracy of Aidoc’s ICH AI tool. They also used that previous study to support its same thesis about AI’s overall generalizability challenges.

New AI Research Chapter – We hear about AI generalizability issues all the time, but post-implementation studies analyzing commercial products are still quite rare, especially studies that specifically name their AI vendor partners.

Two Takeaways – Any AI vendors who find themselves named in critical studies like this (and there will be more), would likely prefer to work with their client to fix any issues, rather than reading about their challenges in a medical journal. Meanwhile, most clinicians who want their AI to work as well as possible would likely view studies like this as an important way to drive necessary improvements. They’d probably both be right.



How AI is Changing Population Health

This Zebra-Med post details how AI is revolutionizing population health programs through automation, workflow integration, and significantly expanding early disease detection.

– Sponsored.



GE’s AI Efficiency & Effectiveness Boost

Learn how GE Healthcare’s growing suite of homegrown and partner AI tools can boost radiology efficiency and effectiveness in its latest post.

– Sponsored.


The Wire

  • Andrew Ng’s Data-Centric AI Campaign: AI leader, Andrew Ng launched a campaign calling for developers to shift their focus to data-centric AI rather than model/algorithm-centric AI, suggesting that the use of mixed/low-quality data is the main reason that models are struggling to generalize in the real world. Ng’s data-centric approach instead calls for AI players to create more systematic ways to clean data and to work with domain experts (e.g. clinicians) to formalize how they assemble and provide data.
  • BWH’s Urban Worklist Update: When Brigham and Women’s Hospital began routing mammograms from its urban mobile van and urban health center to a PACS worklist accessed by all of its breast specialists (vs. a manually-generated paper list provided to a designated radiologist) it significantly improved screening efficiency for these largely minority patients. Exams from the four months before and after implementation (n = 851 & 728) showed a significant turnaround time improvement (101 to 36.4 hours), but statistically unchanged times to diagnostic imaging (39 vs. 45 days) and tissue sampling (43 vs. 59 days).
  • CryptoChart Goes Software-Only: Novarad released a new software-only version of its CryptoChart solution, which allows physicians to share medical images and chart information using secure QR codes (eliminating CDs and/or usernames and passwords). The new CryptoChart version no longer requires a networked router and QR code printer, streamlining setup by allowing providers to use their existing technology.
  • Incidental Breast Cancers on Chest CT: Chest CT isn’t generally known for incidental breast cancer detection, but a recent German study revealed that breast lesions that do turn up on chest CTs are more likely to be malignant, encouraging radiologists to also look for breast lesions when reading chest CTs. The researchers analyzed 35k chest CTs, identifying 31 with incidental breast lesions from 27 patients, including 23 lesions that proved to be malignant (17 carcinomas, 6 metastases).
  • Pulsenmore’s IPO: Israeli home maternity ultrasound startup, Pulsenmore, just completed a $42m IPO (valuing it at $204m) that it will use to fund its commercialization and global expansion. The unique Pulsenmore ES ultrasound system (approved in EU and Israel, seeking FDA approval) allows pregnant women to perform their own fetal scans, using their smartphones to stream exams to their physicians.
  • Modern Vascular Allegations: Outpatient endovascular clinic company, Modern Vascular (14 clinics, 7 states, plenty of IRs on staff), is facing at least four malpractice lawsuits and a U.S. DOJ investigation for allegedly pushing patients to receive unnecessary arterial disease treatment and operating local physician kickback schemes to drive referrals. Nothing has been proven, but there’s more first-hand accounts of Modern Vascular’s questionable operations than we’ve seen in quite some time.
  • LDCT’s Rural Gap: A new ATS Journal study detailed rural Oregonians’ limited access to low-dose CT lung cancer screening, finding that 59% of 29 rural radiology facilities performed LDCT screening and most of these facilities did not follow all “high quality” screening processes. Most of the facilities that did provide LDCT were motivated by community need and the influence of an internal LDCT champion (not financial gain), suggesting that it will require more comprehensive efforts (beyond radiology facilities) to address rural LDCT disparities.
  • Amazon’s DXA Alternative: The Amazon Halo smartphone app can estimate users’ body fat percentage with similar accuracy as DXA-based analysis, using just four smartphone photos from different angles. That’s from an Amazon-backed MGH and LSU study of 134 adults that found the Halo app was similar to DXA-based measurements, and beat all other body fat calculators (e.g. smart scales, air-displacement plethysmography).
  • Predicting Alzheimer’s with MRI ML: A new Radiology Journal study detailed a brain MRI ML model that was able to identify imaging variations in healthy patients and patients with amnestic mild cognitive impairment (aMCI), suggesting that this technique could help spot people with greater risk of Alzheimer’s disease. The model (trained w/ 975 healthy T1 MRIs, tested w/ 270 healthy / 185 aMCI patients) accurately predicted the healthy patients’ age, but overestimated the aMCI patients’ age by an average of 2.7 years.
  • Akumin’s Big Alliance Acquisition: Relatively large U.S. imaging center company, Akumin, is about to get quite a bit larger and more diversified after agreeing to acquire imaging and radiation therapy company Alliance Healthcare Services for $820m. Akumin and Alliance would boast $730m in combined revenue, 154 imaging and 34 radiation therapy locations in 46 states, over 4k team members, and partnerships with over 1,000 hospitals and health systems. That’s a big change from Akumin, which recently had 133 imaging centers in 7 states.
  • Incidental FCPL Follow-Up Variations: A new JACR study detailed wide variations in abdominal radiologists’ follow-up imaging recommendations for small focal cystic pancreatic lesions (FCPLs). The researchers reviewed 2,872 abdominal CT and MRI reports with < 1.5cm incidentally-identified FCPLs, finding 708 reports containing follow-up imaging recommendations (24.7%). Although radiologist recommendations were consistent in emergency and cancer imaging divisions, abdominal radiologists’ follow-up recommendation rates ranged between 25% and 85% of all FCPL cases.
  • Siemens & Prisma: Siemens Healthineers and Prisma Health announced a 10-year value partnership that will provide the South Carolina health system with Siemens’ imaging equipment across all of its sites, an Intelligence Insights Center (AI-supported diagnosis/care and workflow efficiencies), and workforce development support. The Prisma Health value partnership is Siemens’ largest ever in the U.S.
  • Missed Cancer Case Dismissed: A New York appeals court dismissed a malpractice case against a Hudson Valley Radiology Associates radiologist, finding that he satisfactorily interpreted and reported what he believed was a routine screening mammogram based on the information available to him. The now-deceased woman reported a lump in her breast to both the internist who ordered her mammogram and the technologist who performed it, but the internist didn’t specify whether it was a screening or diagnostic mammogram and the technologist didn’t relay information about the lump to the radiologist.

BRG’s PowerScribe One Reporting and Interruption Improvements

When Birmingham Radiological Group-GV adopted Nuance PowerScribe One, the practice eliminated 60-75 minutes in daily reporting time and reduced calls to the radiology reading room by 80% by getting its reports to clinicians faster. See how in this Nuance Case Study.

– Sponsored.


Ever wonder where The Imaging Wire finds its news?

One of our favorite sources is the AI for Radiology newsletter.

The folks producing AI for Radiology are academics, so there’s no commercial motives or clickbait titles. Instead, they leverage their deep understanding of imaging AI and strong industry connections to fill each monthly issue with inside scoops and expert perspectives.

Whether you’re on the clinical or commercial side of imaging AI, this mighty little monthly newsletter is worth a spot in your inbox.


The Resource Wire

  • A new study in European Radiology highlighted Riverain Technologies’ ClearRead Xray – Detect as one of just two imaging AI products to achieve the FDA’s most stringent premarket approval level. See how they measured up against the other 99 AI tools here.
  • In this Novarad video, interventional oncologist Gary M. Onik, MD shares how Novarad’s AR surgical navigation system, OpenSight, helps his team accurately assess and treat tumors.
  • Check out this UCSD lung nodule detection study detailing how Arterys Lung AI drove a “clinically meaningful and statistically significant increase in sensitivity,” without changing reading time.
  • Independent and staying that way? Healthcare Administrative Partners just released a helpful set of guidelines that radiology practices can follow to stay private despite ongoing consolidation pressures.

Today’s issue was brought to you by Jake Fishman and Jason Barry.

You're signed up!

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

-- The Imaging Wire team