COVID Shortcuts | AI Stewards | Superficial US

“Don’t want to use AI? Tough.”

The PACSman, Michael J. Cannavo, with a reminder that using imaging AI might be unavoidable.

Imaging Wire Sponsors

  • Bayer Radiology – Providing a portfolio of radiology products, solutions, and services that enable radiologists to get the clear answers they need.
  • GE Healthcare – Enabling clinicians to make faster, more informed decisions through intelligent devices, data analytics, applications and services.
  • Healthcare Administrative Partners – Empowering radiology groups through expert revenue cycle management, clinical analytics, practice support, and specialized coding.
  • Hitachi Healthcare Americas – Delivering best in class medical imaging technologies and value-based reporting.
  • Nuance – AI and cloud-powered technology solutions to help radiologists stay focused, move quickly, and work smarter.
  • Riverain Technologies – Offering artificial intelligence tools dedicated to the early, efficient detection of lung disease.

The Imaging Wire

COVID AI Shortcuts

A new preprint study out of the University of Washington found that even though some recent CXR AI models have proven to accurately detect COVID-19 in the lab, they may actually rely on confounding factors to support their results (not medical pathology) making them unlikely to perform well in the real world. Here are some details:

The Theory – The team observed that many recent models intended to detect COVID-19 in chest X-rays followed a “near worst-case” training method, using one data source for COVID-negative images and a separate source for COVID-positive images. Because of this, the models are more likely to interpret systematic differences between the sources as evidence of a patient’s COVID status.

The Study – To test this theory, the researchers created two datasets. Dataset I used separate sources for CV19-positive (GitHub-COVID, n = 408) and negative (NIH ChestX-ray, n = 112k) patient images, while Dataset II used positive and negative images from the same healthcare system (n = 1.6k positive, 96k negative). The team then trained models using each dataset, tested them against both internal and external sets, and ran a series of experiments.

The Results – The team found that although their algorithms appeared accurate when tested against internal data (Dataset I: 0.992 AUC, DS II: 0.995 AUC), accuracy fell dramatically when tested against the external datasets (Dataset I: 0.76 AUC, DS II: 0.70 AUC). Using saliency maps, they found that confounding factors such as patient positioning and additional image features (laterality markers, annotations) played a major role in the external tests’ AUC declines.

Takeaways – The study’s main takeaways are that AI will take shortcuts to reach a conclusion, so CXR-based COVID-19 algorithms should be validated with multi-site external data and investigated using AI explanation methods (saliency maps/GANs). AI shortcuts and the need for external validation might be a given for AI insiders, but these are still important reminders amid the rush to fight COVID with AI.

The Case for Algorithmic Stewardship

A new JAMA paper encouraged hospitals to adopt the practice of AI/ML stewardship, applying components of the 200-year-old practice of hospital stewardship to modern medicine. Much like other current stewardship programs (e.g. antimicrobial or pharmaceutical stewardships), the authors suggest that algorithmic stewardship could help ensure AI quality for hospitals that would otherwise struggle to manage their AI practices. Here’s how:

AI Inventory – The paper called on health systems to maintain an inventory of all predictive algorithms that they use, placing an emphasis on the predicted outcomes and the diagnostic decisions that they influence.

AI Audits – The authors suggest that AI stewards should audit algorithms for safety and fairness across diverse patient populations before they are brought into use or upon major changes (e.g. new / updated algorithms, patient population changes).

AI Monitoring – Algorithmic stewardship programs would also conduct routine evaluations of AI performance in actual clinical settings, reviewing how data or algorithm changes (e.g. EHR inputs, learned imaging shortcuts) might influence model performance.

AI Adaptation – Like other stewardship practices, AI programs should aim to identify and adapt existing best practices that are already working well within their institutions to improve overall AI operations.

The Wire

  • Clarius Gets Superficial: Clarius Mobile Health announced the launch of the Clarius L20, positioning it as “the world’s first ultra-high frequency handheld ultrasound,” and highlighting its ability to perform superficial scanning (skin to 4cm, 8-20MHz frequency) combined with its low $6,900 price point. The Clarius L20 HD will target clinical settings that currently use ~$35k cart-based systems to visualize shallow anatomy (e.g. MSK, rheumatology, podiatry, plastic surgery, microsurgery, pediatric anesthesia).
  • FDA AI Database: A new paper in NPJ Digital Medicine detailed the 64 medical AI devices and algorithms cleared by the FDA through February 2020 (46.9% radiology-focused), while promoting the authors’ online database of FDA-approved AI/ML-based medical technologies. The paper’s infographic and online database are its most eye-catching sections, but the authors’ main goal is to show that regulators technically could define medical products as AI/ML-based and maintain their own AI/ML databases if they decided to do so (the FDA doesn’t do this).
  • Jubilant Dismissal: An Alabama federal magistrate recommended the dismissal of antitrust allegations against Jubilant Radiopharma, noting that the 2019 monopoly lawsuit failed to prove that any competitors were “willing and able” to offer the two radiopharmaceuticals that Jubilant allegedly has a monopoly over.
  • AI Economics: The one and only PACSman (aka Michael J. Cannavo) shared a crash course in imaging AI economics on Auntminnie.com this week, prompted by Viz.ai’s new technology add-on payment (NTAP). There’s a lot to this editorial, but PACSman’s big takeaways are: 1) AI reimbursements could bring major changes to the business of AI; 2) AI companies should focus more on market education; 3) VC firms need to extend their AI startup ROI timelines; 4) AI firms should become more customer experience-focused (vs. tech-focused).
  • MRI for COVID Follow-Ups: A new paper in Clinical Imaging detailed COVID-19 pneumonia features in chest MRI scans, suggesting that chest MRI could serve as an alternative to chest CT for COVID-19 follow-up scans, particularly when ionizing radiation exposure is a concern. The researchers performed chest CTs and MRIs on eight COVID-positive patients, finding that the same CV19 features were visible in both the CT and MRI scans (5 w/ typical features, 3 w/ atypical).
  • Covera’s PSO Cert: Radiology clinical analytics firm, Covera Health, announced its certification as a patient safety organization (PSO) by the Agency for Healthcare Research and Quality (AHRQ). Given that Covera certifies radiology practices for their diagnostic accuracy, the PSO certification certainly aligns with its core focus.
  • RSIP Knee Replacement Module: RSIP Vision rolled out a new AI bone segmentation and landmark detection module that uses X-rays to streamline pre-operative planning and intraoperative guidance for knee replacement surgeries. RSIP Vision’s new X-ray knee replacement module joins the company’s similar CT-based module, making the solution available to the many medical centers who rely on X-ray for these procedures.
  • PET/CT CNN for Lymphoma: A University of Wisconsin team developed a CNN that analyzes 18F-FDG PET/CT scans to automatically detect lymph nodes with lymphoma involvement. In a retrospective study of 90 adult patients with lymphoma who received 18F-FDG PET/CT scans, the algorithm was able to identify affected lymph nodes with an 85% true positive rate (923 of 1087 nodes). Among the 20 patient scans that were also reviewed by two nuclear medicine physicians, the algorithm’s 90% true positive rate (197 of 219 nodes) was “nearly comparable” to the second reader’s 96% rate.
  • China’s Image Database: The Chinese government began work on a national medical image database, which could significantly expand how Chinese institutions store (currently only in hospitals) and share (currently only shared among affiliated hospitals) medical images. China also intends to use the national image database to advance the country’s imaging AI development leadership. The Chinese Society of Radiology (CSR) will oversee the creation of the standardized medical image platform, collaborating with between 350 and 400 participating hospitals.
  • RADLogics’ Pneumothorax AI: RADLogics announced the FDA clearance of its chest X-ray pneumothorax application for patient triage and prioritization. The new solution flags CXR scans with suspected pneumothorax (collapsed lung), making case-level output available to a PACS workstation for worklist prioritization or triage.
  • No-Cost CAC Works: A new study by University Hospitals (Cleveland, OH) found that eliminating out-of-pocket costs for coronary artery calcium screening (CAC, usually $400-$800) drove an immediate increase in CAC utilization. The study looked at 27,466 patients, including 5,109 “low-charge” patients (paid $99) and 22,357 “no-charge” patients (paid $0), finding that the no-charge model drove a 546% increase in monthly CAC utilization, while improving CAC utilization among older patients (57.9 vs 59.3 years), women (46% vs 51% share), and black patients (7.2% vs 9.4% share).
  • Philips Azurion Lung Edition: Philips unveiled its new Azurion Lung Edition 3D imaging and navigation platform for lung cancer diagnosis and treatment. The suite combines tableside Cone Beam CT scans with live X-ray guidance and advanced tools to support image-guided lung procedures, allowing clinicians to perform endobronchial biopsies and lesion ablation during the same procedure.

The Resource Wire

– This is sponsored content.

  • This GE Healthcare Insight post details how improving the radiology workflow with AI are improving radiology workflows across a range of modalities and clinical roles.
  • In Nuance’s latest Q&A, Diagnostics leader, Karen Holzberger, sat down with Dr. Irena Tocino from Yale New Haven Health System to learn about how Nuance solutions helped YNHHS overcome the challenges brought by the COVID-19 pandemic.
  • This Hitachi Healthcare Americas’ blog details COVID-19’s recent and future impact, warning cardiac practices and clinics of an upcoming wave of patients with cardiovascular issues that worsened due to delayed treatments, followed by a “third wave” of patients who developed heart complications from COVID-19 infections. Hitachi also shared some guidance on how to manage and minimize these waves.
  • This Riverain Technologies case study details how Duke University Medical Center integrated ClearRead CT into its chest CT workflows, reducing read times by 26% and improving nodule detection by 29%.

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!