Uncategorized

Google’s Breast Cancer Model | Quality Consolidation

“It feels like this is another step towards this technology actually making a difference in the real world.”

Google Health’s U.K. lead, Dominic King, after announcing study results on the company’s new breast cancer screening algorithm.



We’re thrilled to announce that GE Healthcare is now an Imaging Wire sponsor. GE Healthcare’s point of care ultrasound team does awesome work and we’re excited to share their message with you. Every issue of The Imaging Wire is made possible by GE Healthcare and the rest of our sponsors, so make sure to keep up with the great things they’re doing.


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
  • GE Healthcare – Providing point of care ultrasound systems, from pocket-sized to portable consoles, designed to support your clinical needs and grow along with your practice.
  • 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
  • Qure.ai – Making healthcare more accessible by applying deep learning to radiology imaging

The Imaging Wire

Google’s Breast Cancer Model

Google was the talk of the radiology industry last week after introducing a breast cancer detection AI model that can reportedly match or beat radiologists. Here are the details:

Training & Testing – The Google Health model was trained on two mammogram datasets from women in the U.K. and U.S. (n = over 76k & 15K) and designed to spot signs of breast cancer. The models were then evaluated using test sets from the U.K. (n = 25,856 women, 785 biopsied, 414 diagnosed w/ cancer within 39 months) and the U.S. (n = 3,097 women, using images of all 1,511 women who were biopsied and random sample of unbiopsied women, 686 diagnosed w/ cancer within 27 months).

Results Highlights – There’s a lot to these studies, but the big takeaway is the Google model was able to spot cancers that were originally missed by the radiologists (reduced false negatives by 9.4% in U.S. and 2.7% in U.K.) and reduced false-positives (cut false positives by 5.7% in U.S. and 1.2% in U.K.). The model achieved a respectable 0.899 AUC in the U.K. and less-impressive 0.757 AUC in the U.S.

U.K. Time Savings – Google’s retrospective study in the U.K. found that the model improved first reader specificity by 1.2% and sensitivity by 2.7%, while proving to be non-inferior to second reader radiologists. That last part is key, as matching the second reader has the potential to reduce mammography reading volumes by 88% in countries that commonly use a second reader for breast cancer screening.

U.S. AUC Improvement – The U.S. study compared the model to BI-RADS findings from a panel of six radiologists (n = test set of 500 mammograms plus available histories, 125 with cancer), achieving an 11.5% higher AUC-ROC than the average radiologist. The radiologists and model were even better together, as sometimes all six radiologists caught cancers that the model missed and the model identified cancers on other scans that each of the radiologists missed.

Generalization Test – In an effort to test if the model could generalize to other healthcare systems or regions, Google Health then trained the model on data from the women in the U.K. and then evaluated it on a U.S. dataset, achieving a 3.5% reduction in false positives and an 8.1% reduction in false negatives. Google didn’t reveal if the scanner technology was also different.

Significance – Although Google’s focus on medical image interpretation is well known and the company has already developed a number of other healthcare algorithms (lung cancer, eye disease, and kidney injury), this study generated about as much radiology AI buzz as anything in recent memory. Folks in the AI community had a generally positive take on Google Health’s latest model, partially because they viewed the methodology as solid (though not perfect), and partially because it’s a sign that healthcare AI is growing up. Meanwhile, Google’s reputation and marketing clout brought the science and potential of imaging AI (not just the hype/concerns) to the TV and PC screens of millions of people in a way that’s rarely been done before.

Next Up – Google noted that the model would require “continued research, prospective clinical studies and regulatory approval” in order to make it to clinical use (and was vague about whether it plans to do these steps), but did reveal that it’s working to understand the best ways to incorporate the model into clinical workflows.



Quality Consolidation

A new study in NEJM revealed that the quality of care at acquired hospitals generally got worse or stayed the same after acquisition, dealing a blow to one of the main rationales for hospital consolidation. Here are the details from the first major study on consolidation’s impact on care:

The Study – The Harvard study of 246 hospitals acquired between 2009 and 2013, plus nearly 2,000 control group hospitals, compared four quality of care measures before and after acquisitions (clinical processes, patient experience, mortality, readmission rate).

The Results – The study found that acquired hospitals saw a modest decline in patient experience (from 50th to the 41st percentile) and no change in 30-day readmission and 30-day mortality rates. The study also found a notable improvement in clinical processes (0.05 to 0.38) after acquisitions, but the researchers didn’t attribute the process improvement to the acquisitions.

Significance – These results could be worse, but they certainly don’t support larger healthcare systems’ argument that hospital M&A results in improved patient care or justify the higher healthcare costs that are known to come with hospital consolidation. The study could also give regulators new reason to push back against hospital acquisition proposals that lead with promises of improved care.


The Wire

  • Check-Cap announces positive results from its C-Scan capsule-based X-ray colorectal cancer screening system’s U.S. pilot study, revealing that it is advancing towards FDA submission and a U.S. pivotal study in late 2020. The pilot C-Scan study (n= 28 patients, prospective, multi-center, open label, single arm) did not have any serious adverse events, achieved higher patient satisfaction compared to colonoscopy (no surprise there), and its patient results revealed an agreement between C-Scan and colonoscopy for polyp detection.
  • 24x7mag.com’s recent HTM Salary and Satisfaction Survey (n = 1,438 HTMs) revealed that nearly all U.S. Healthcare Technology Management salaries rose in 2019, with radiology equipment specialists achieving the greatest growth (up from $85k to $92k) and HTM management roles retaining the highest salaries ($120k to $125k). The survey also revealed a looming HTM talent shortage due to expected retirements and insufficient HTM training programs, noting that there are nearly 2-times more HTMs over 55 years-old than HTMs under 35 years-old in the workforce.
  • New research in the Journal of Hepatology provided solid evidence in favor of using non-enhanced MRI for hepatocellular carcinoma surveillance among high-risk patients (vs. ultrasound). The study (n = 382 high-risk patients, 43 w/ HCCs) found that MRI had significantly higher per-exam sensitivity (79.1% vs. 27.9%), specificity (97.9% vs. 94.5%), positive predictive value (61.8% vs. 17.7%), and negative predictive value (99.1% vs. 96.9%) than ultrasound, while also allowing scans in under 6 minutes and avoiding any contrast agent concerns. MRI surveillance’s main drawback was its 25 to 35-minute required patient stay for each scan compared to ultrasound’s 12 to 15-minute stay.
  • A recent ABC News Australia story detailed the lack of modern, wide-gantry MRI scanners that are available to the country’s Medicare patients, blaming Australia’s Medicare licensing system that typically only covers scans from older MRI machines (no other modality requires a government license for Medicare rebates). As a result, Australian Medicare patients who require large-gantry MRI scans are waiting up to six months, traveling for hours, or paying out-of-pocket, while the country’s MRI fleet experiences unbalanced demand (e.g. new & unlicensed MRIs are sitting idle).
  • New research from a University of Michigan team explored their outpatient imaging exam callbacks between October 2015 and October 2017, discovering that callbacks are rare but often have common sources/causes that could be addressed. MRI callbacks (0.114%, 168 of 147K exams) occurred at a 9-times higher rate than CT callbacks (0.013%, 26 of 195.5k exams) due in part to high callback rates for MSK exams (MRI: 0.265%, 73 of 27.5k exams; CT: 0.090%, 7 of 7.8k exams). The vast majority of callbacks were driven by specific radiologists (top 9 of 65 rads initiated 52% of callbacks, top-20 initiated 80%) and were due to specific preventable issues (protocol errors 28%, inadequate anatomic coverage 21%, incomplete exams 13%, perceived suboptimal image quality 11%).
  • OTech shared its top ten imaging informatics trends of 2020, following the Gartner hype cycle format to show where each trend is positioned and where it’s headed. OTech identified AI, VR, and Blockchain as its key early hype cycle trends, while positioning POC ultrasound, 3D printing, and FHIR higher up on the “Inflated Expectations” arc. The Cloud and the Internet of Medical Things (IOMT) are positioned beyond the hype peak and heading downward towards the “Trough of Disillusionment,” while VNA and Digital Radiography have graduated and are headed on to the “Plateau of Productivity.”
  • A UCSF and UC Berkeley longitudinal study (n= 32 patients with early Alzheimer’s) found that Tau PET is able to predict the location of Alzheimer’s brain atrophy 15 months in the future (β-amyloid–PET couldn’t), potentially supporting the development of new treatments. The study found that Tau PET signal distribution was a strong indicator of future atrophy at the group and patient levels (predicting ~40% of future atrophy), specifically among younger patients (~60% of future atrophy).
  • New Mexico’s Roosevelt General Hospital warned its patients to monitor their credit scores following a potential theft of their personal information from a malware-infected digital imaging server (including: names, addresses, birth dates, drivers license numbers, SSNs, phone numbers, medical info, and gender). It is unclear whether any patient information was actually stolen, but the hospital is providing monitoring services to the patients as a precaution.
  • New research from a New York Presbyterian Hospital team found that referring physicians comply with radiologist recommendations from musculoskeletal MRI exams at a relatively high 73% rate, while recommendations for additional imaging was the most common recommendation type (66%) and had the lowest compliance rate (63%). The study suggested that follow up compliance for additional imaging recommendations could be improved through better patient communication (so they understand why it’s necessary) and more direct communication with other clinicians.
  • Hologic agreed to sell its struggling Cynosure medical aesthetics business to PE firm Clayton, Dubilier & Rice for $205 million (receiving $138m after closing adjustments) and revealed plans to record a $155 million to $185 million impairment charge as a result of the deal. Cynosure “significantly underperformed” expectations since Hologic acquired it in 2017 for $1.6 billion, although the company reassured that it’s still committed to making smaller, tuck-in acquisitions that strengthen its core franchises. Hologic also announced plans to buy back $205 million in common stock in addition to the company’s existing $211 million repurchase plan.
  • A recent AJR study (n= 2,411 patients with suspected pneumonia) revealed that chest radiography has a much higher rate of uncertain radiology report impressions for pneumonia diagnosis than CT (31.8% vs. 21.7%) and detailed how uncertain/ambiguous impressions were more likely to result in poor sensitivity/specificity with X-ray-based reports than CT-based reports. Given that most previous studies on this subject excluded uncertain impressions, the researchers suggest that chest radiography may be less valuable for pneumonia diagnosis than previously believed.
  • The ACR highlighted how the U.S. government’s 2020 spending bill will bring millions in funding for radiology-related projects and agencies including funding for the NIH ($41.7 billion), Agency for Healthcare Research and Quality ($338m), precision medicine research ($500m), Brain Research through Neurotechnologies ($500m), Cancer Moonshot ($195m), and the Childhood Cancer Data Initiative ($50m). The ACR also announced that Congress’ appropriations packages included a provision to restore and extend pass-through payment status for two radiopharmaceutical drugs used in the “Imaging Dementia—Evidence for Amyloid Scanning (IDEAS)” Study (the drugs’ status previously expired in 2017 and 2018).

The Resource Wire

  • The GE Healthcare Venue Go combines a uniquely adaptable design with AI-enabled tools to support fast triage and help you make confident diagnoses and informed decisions.
  • In this Nuance video, Penn Medicine professor, Warren B. Gefter, shared how PowerScribe One leverages AI, structured data, and automation to drive improved patient care.
  • By partnering with Medmo, imaging centers can keep their schedules full, their equipment busy, and increase revenue. Here’s where to get started.
  • The first patients were recently treated in a clinical trial using focused ultrasound to enhance the effectiveness of chemotherapy drugs in those with Her2+ breast cancer that has metastasized to the brain.

You're signed up!

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

-- The Imaging Wire team