“I went into it with some worries, but this exceeded my fears.”
Maastricht University’s Laure Wynants after evaluating hundreds of faulty COVID AI models.
Imaging Wire Sponsors
Arterys | Bayer Radiology | Blackford Analysis
Canon Medical Systems | Fujifilm Healthcare Americas
GE Healthcare | Novarad | Nuance
Riverain Technologies | Siemens Healthineers
United Imaging | Zebra Medical Vision
The Imaging Wire
A new study with some high-profile authors revealed that imaging AI models can accurately predict patient race using images that don’t even include racial identifiers. That’s interesting on its own, but it also suggests that AI models carry race-based diagnostic risks that human radiologists won’t be able to catch.
- The (pre-print) Study – The researchers developed deep learning models using multiple CT and X-ray datasets and evaluated: A) Imaging AI’s ability to detect patient race, B) Possible confounding features that might cause AI to identify races (e.g. disease distribution, body mass), and C) How and why AI models can do this.
- Accurate Predictions – The AI models VERY accurately predicted patients’ self-identified race (all tests >0.80 AUC, most tests >0.90 AUC, some tests 0.99 AUC). In fact, the models accurately identified race with images from multiple modalities, clinical tasks, and quality levels (e.g. corrupted or cropped images) – even with models that were made for clinical use.
- No Proxies or Confounding Features – The researchers confirmed that these racial predictions were “not due to trivial proxies” or imaging-related confounding features. Yes, they “double checked.” However, they weren’t able to figure out why AI is so good at detecting race in X-rays and CTs. They also believe it will be difficult to fix this issue.
- Viral Research – It’s safe to say that this study set Radiology / AI Twitter ablaze (604 retweets, 178 comments, 670 likes), given its combination of hot-button issues (race, healthcare disparities, explainable AI) and being featured in the latest blog from Dr. Luke Oakden-Rayner (he’s still a must-read). It’s also very safe to say that not everyone on Twitter was comfortable with this study or its interpretations.
- The Takeaway – Imaging AI models intended for non-racial clinical tasks can apparently identify patients’ races and therefore could be applying the same unknown racial “biomarkers” when it creates its diagnoses. It’s going to require more research to understand how this is happening and it’s going to place even more pressure on commercial AI players to validate their performance across different racial groups.
Blackford on Pediatric AI
Check out this Blackford Analysis white paper detailing how children’s hospital imaging teams can leverage AI to improve modality throughput and imaging device availability.
Why United Imaging’s CT (uCT)?
United Imaging’s Z-Detector architecture enables high resolution and low noise imaging across the entire uCT portfolio. United Imaging also offers the same CT quality and software platform across all systems (you choose your slices and speed), and all uCTs can image a wide variety of patient and exam types, including large patients, patients with metal implants, angiography, and cardiology exams.
- CMS Renews Viz.ai LVO: Viz.ai announced that CMS renewed its Viz LVO stroke detection software’s highly celebrated NTAP through FY2022 (reimbursing providers up to $1,040 for each qualified use), maintaining that Viz LVO still saves time, improves outcomes, and increases access to care. It will be interesting to see if, like last year, competing stroke AI companies also announce their continued NTAP qualification because they are “substantially similar” to Viz LVO.
- What Went Wrong with COVID AI? MIT Technology Review detailed how unhelpful hundreds of COVID AI models proved to be, blaming these results on a long list of training and testing mistakes (redundant/siloed AI development efforts, poor data / data management, lots of bias sources, poor clinician/developer coordination). On the bright side, the article suggests that learning from these mistakes will help healthcare AI improve going forward.
- The Case for Direct-to-Angio: A new JAMA Neurology study supported bringing patients with acute LVO strokes directly to the angiography suite (without stopping for CT). The randomized trial (n = 174 w/ suspected LVO) found that direct-to-angio patients were more likely to undergo endovascular treatment (83.5% vs. 71.9%), had shorter median puncture and reperfusion times (18 vs. 42 mins & 57 vs. 84 mins), and had better outcomes. This comes one week after another study highlighted using CB-CT in the angio suite (also skipping conventional CT).
- Blackford Adds Qlarity: Qlarity’s QuantX Diagnostic AI breast MR decision support tool is now available on Blackford Analysis’ curated AI marketplace. The addition gives the Blackford marketplace a unique breast imaging AI tool (MRI-based, focused on high-risk women), while improving Qlarity’s visibility and integration process among Blackford platform users.
- Post Shutdown Cancellations: A new Brigham and Women’s/Harvard study revealed that mammography screening cancelation rates increased after the COVID shutdown (46% vs. 37% before), with the highest cancellation risks among non-white patients (1.34 vs 1.25 white), 70yr-old women (1.36 vs. 1.20 53yrs), and Medicare patients (1.41 vs. 1.26 Medicaid and 1.21 “other”).
- Open-Source AI: Stanford’s AIMI center took a big step towards democratizing AI research, launching a new open-source image data and training platform that it hopes will eliminate academia’s AI development cost/labor/resource barriers. The Microsoft-supported platform will start with over 1m annotated medical images, and will later add more images from Stanford and other organizations, as well as a range of AI development resources (computing power, standardized ML tools, pre-trained models).
- Cervical CT Spike: A new Emory study found that emergency cervical spine imaging increased at a 5.7% annual rate between 2009 to 2018, driven in part by scans for minor injuries (+10.9%/yr), and resulting in a significant increase in cervical CT scans (+10.5%yr, +136% overall).
- Radiographer Cross-Training: A Hospital Times UK editorial called on the NHS to upskill its radiographers to help the country tackle its cancer treatment backlog. The editorial suggests that radiographers should be trained to support a range of radiotherapy duties, including assisting with radiation therapy procedures (e.g. SABR) and offloading certain radiation oncologist tasks (e.g. contouring). The editorial didn’t address how this would impact the NHS’ cancer imaging backlog.
- RealView’s Holographic FDA: RealView Imaging announced the FDA approval of its HOLOSCOPE-i Holographic System, which allows physicians to visualize 3D medical holograms based on CT and ultrasound data before/during interventional procedures. The HOLOSCOPE-I projects 3D holograms in front of the physician (no glasses etc.), which they suggest allows prolonged use without the problems associated with wearable AR/VR systems (fatigue, nausea, headache).
- Scanxiety Prevalence: A new Supportive Care in Cancer study (n = 222) revealed that 55% of patients with advanced cancer experience scan-associated anxiety (‘scanxiety’), with the highest scanxiety rates among patients who are female, rural, have breast cancer, or have depression or anxiety (65%, 66%, 68%, 73%, 89%). Scanxiety severity levels (1-10 scale, 6 avg.) were highest among patients diagnosed within the last year, patients with result wait times under 2 days, and patients with anxiety (6.7, 6.7, 7.2).
- Wearable X-Ray: A team of US and China-based scientists developed a flexible X-ray detector that could eventually lead to wearable X-ray detectors for radiation monitoring and medical imaging. The proof-of-concept detector owes its wearability to its use of metal-organic frameworks (no lead or heavy metals) and flexible electrodes (not stiff). Initial studies show that the wearable X-ray is more sensitive than lead-based detectors and can sustain heavy use.
- Pulse’s Big Series C: Chinese coronary imaging company, Pulse Medical Imaging Technology, just landed a massive $100m series C round that it will use to fund R&D, clinical trials, and global expansion. Pulse might not be a household name in the West, but it has some interesting technology (measures FFR via angiography, OCT, intravascular ultrasound exams), and its Series C round included funding from Philips.
- ED Neuroimaging Growth: A new AJR study detailed emergency neuroimaging utilization’s “considerable growth” between 2007 and 2017 (+72% overall, +5%/yr), driven by head and neck CTA increases (+ 1100% & 1300% overall) and still dominated by non-contrast head CTs (+69% overall). The authors attributed these increases to a lack of consensus on emergency neuroimaging guidelines, calling for more research into neurovascular imaging appropriateness as well as overall ED imaging.
Why Adopt Contrast Dose Management
With radiation dose management now largely considered best practice, this Bayer white paper details the top five benefits of adopting contrast dose management.
The Resource Wire
- Watch CriticalCare’s Dr. Yusuf Karrar, MD discuss how Canon’s CT Fluoro touch interface allows “Simple, Streamlined CT Fluoroscopy.”
- See how UCSD reduced its lung nodule detection times and increased its reader performance using Arterys Lung AI.
- Tune in to Nuance’s AI webinar (today at 2pm Eastern), where they’ll detail how AI is supporting radiologists, and how AI outputs are informing providers and specialists across the care continuum.
- Watch Jared Christensen, MD, MBA explain how Duke University Health uses Riverain Technology’s ClearRead CT Vessel Suppress and ClearRead CT Detect in its daily practice.
- Room for more efficiency in your breast imaging operations? Check out this GE Healthcare post detailing how new technologies are improving patient experiences and making breast imaging teams more efficient.
- See how New Jersey’s Ramapo Radiology Associates overcame their CD burning problems and improved their physician and patient experiences with Novarad CryptoChart.
- See why the time is right for imaging AI-enabled population health in this Hardian Health profile featuring Zebra-Med’s CEO, Zohar Elhanani.
- This Diagnostic and Interventional Cardiology article details the unique advantages of cloud-based CVIS systems (off-property access, team collaboration), with insights from one Mississippi-based cardiologist on the benefits of Fujifilm Healthcare’s VidiStar CVIS.
- Be at the forefront of PET/CT with an extended field of view. Download this Siemens Healthineers clinical case gallery to see how the extended field of view on Biograph Vision Quadra could reshape clinical outcomes.