LDCT Screening Gap | Toy AI | MRI Moonshot


“I think it’s an unstoppable train . . .”

Harvard Medical School professor, Isaac Kohane’s big expectations for imaging AI.




Imaging Wire Sponsors

  • Arterys – Transforming medical imaging by reducing the tedium of interpretation and empowering physicians with our web-based AI platform.
  • 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.
  • Siemens Healthineers – Shaping the digital transformation of imaging to improve patient care.



The Imaging Wire


LDCT Screening Works, and Needs Work

Last week brought even more evidence that low-dose CT lung cancer screening saves lives, but also revealed that screening programs have a long way to go before they are saving all the lives they could. Here are some details:

  • LDCT Screening Works – A University of Georgia study confirmed the “significant” lifesaving benefits of low-dose CT lung cancer screening, “albeit with a tradeoff of likely overdiagnosis.” The study analyzed eight trials (n = 90,475 patients) finding that patients who participated in LDCT screening had a 0.4% lower risk of lung cancer death (1.72% vs. 2.12%), essentially avoiding one cancer death for every 250 participants (better than breast cancer screening). Screening also had a 20% higher overdiagnosis rate (high, but still similar to breast cancer screening).
  • LDCT Screening Programs Need Work – However, LDCT screening adoption in the U.S. has been slow and uneven. National Cancer institute researchers found that screening rates among eligible adults slowly increased from 2016 to 2018 (3.3%, 3.4%, 5%), leaving the other 95% of eligible adults unscreened. The study also revealed serious regional discrepancies, as three southern states with the highest lung-cancer burden had low screening rates (MS, WV, AR = <4%), while northern states with the lowest lung cancer burden had the highest screening rates (MA, VT, NH = 12.8–15.2%).



Sight & Sound AI

A new Radiology: Artificial Intelligence study detailed an algorithm that accurately labels brain MRI images using eye-tracking and speech recognition software, potentially streamlining the AI training process.

  • Sight & Sound Training – The researchers developed the CNN algorithm using recordings of a radiologist’s dictation and eye focus while interpreting 700 two-dimensional brain tumor MRI scans.
  • Simulated Interpretation – They used this model to automatically label lesion locations in an 85-image test set with 92% accuracy, “demonstrating that fully simulated interpretation can yield reliable tumor location labels.”
  • Simulated Application – The researchers then used the automatically-labeled images to train another model that predicted lesion location on a separate testing set with 85% accuracy.
  • Significance – We’re still in the very early phases of this type of AI development, but this is a really interesting example of how to combine medical imaging, NLP, and eye tracking for human-in-the-loop AI development.

The Wire

  • Toy AI: A new paper in European Radiology criticized the rising number of COVID AI algorithms, suggesting that many are trained on “ill-curated . . . toy datasets,” developed solely by AI researchers (no radiologists involved), hastily published without peer review, and often don’t have clinical relevance (since imaging isn’t approved for COVID diagnosis). Although these “toy” algorithms are unlikely to make it to clinical use, the authors warned that they may still cause harm by creating false hope among patients and misguide less experienced scientists and radiologists.
  • Caption’s LUS Grant: Caption Health landed a $4.95m grant from the Bill & Melinda Gates Foundation to create lung ultrasound guidance tools that would help less experienced clinicians diagnose pneumonia. This solution would expand Caption Health beyond cardiac ultrasound guidance and help address lung ultrasound’s user expertise challenges.
  • COVID CT Variations: A new study in the Radiology journal detailed wide variations in CT utilization, protocols, and radiation doses for patients with COVID-19 pneumonia, calling for increased guidance. The study (n = 782 patients, 54 sites, 28 countries) found that radiation exposure varied widely (7 to 11mGy), the majority of sites use CT to assess severity (76%) but many sites still use CT for initial diagnosis (22%), and single-phase CT is far more prevalent than multi-phase (80% vs. 20% of sites).
  • History Helps: When radiologists are provided patients’ clinical history, their diagnoses improve – most of the time. That’s from a review of 22 studies, including 15 that showed clinical history improved diagnostic accuracy, 6 studies that found it didn’t change accuracy, and one study that found a decline in diagnostic performance.
  • Canon & Avicenna.AI’s Stroke Partnership: French AI developer Avicenna.AI announced that Canon Medical will integrate its CINA Head stroke triage solution into the Canon Automation Platform. CINA Head (FDA approved) triages CT images for intracranial hemorrhages and large vessel occlusions in emergency room settings, automatically detecting and prioritizing the conditions within 20 seconds and with 96% accuracy.
  • Structured COVID CXR Predictions: A new study out of Ireland detailed a structured CXR reporting system that was able to predict COVID-19 diagnosis with high accuracy and interrater reliability. The study had two blinded radiologists review CXRs from 582 patients (143 CV19-positive) and classify them based on a structured reporting template (Characteristic, High Suspicion, Indeterminate, Unlikely and Normal), correctly predicting 88% of the patients’ RT-PCR results and achieving a good interrater reliability (Cohen’s κ = 0.8).
  • Vanderbilt’s MRI Moonshot: A Vanderbilt team will use a $1.4m NIH grant to develop a 47.5 millitesla MRI that’s smaller, lighter, silent, less expensive, and more portable than current systems. To make this happen, the team will develop new hardware (low-field radio frequency transmission coils, amplifiers) and software, and use spatial encoding approaches that haven’t been used in clinical MRIs.
  • DL Density Predictions: A Whiterabbit AI and Washington University-led study published in Radiology: Artificial Intelligence detailed a deep learning model that accurately predicted BI-RADS breast density. Team trained the model using 57.4k digital mammography images and evaluated it using synthetic mammography datasets from two institutions (n = 14.4k and 63.9k images), achieving “substantial agreement” with the exams’ original density findings. They then used an adapted version of the model at one of the sites (trained with 500 SM images), improving the site’s BI-RADS breast density performance.
  • The Case For and Against ABR MRI: A pair of AJR e-Views share conflicting perspectives on abbreviated MRI protocols (specifically MSK). The case for ABR-MRI seems pretty straightforward: its shorter exam times improve efficiency, access, and costs. However, the counterpoint suggested that abbreviated MRIs might actually hurt care (can hinder diagnostic confidence, not appropriate for some cases) and exacerbate radiology’s burnout problems.
  • The COVID Lockdown’s Invisible Costs: A new BMJ article detailed how the COVID-19 lockdowns created what they call a cancer pandemic, noting a drop in patients seeking diagnosis, referrals from primary care, and delayed diagnoses. The authors cited a related UK study finding that the COVID-19 pandemic will lead to 3,500 avoidable cancer deaths (60k lost years), warning that treatment delays would further increase this number.
  • AI Revolution: The Harvard Gazette published a super detailed perspective on medicine’s upcoming AI revolution, reviewing healthcare AI’s many expected benefits (improve diagnosis / care / efficiency / costs / access) and warning of its “equally big” challenges (misdiagnoses, bias, how it will interact with unpredictable humans). In order to achieve these benefits and avoid these challenges, the article suggests that developers create AI with these potential pitfalls in mind, leverage insights from a range of outside disciplines for AI design (e.g. ethics and philosophy), perform real-world testing, and regularly re-evaluate performance.
  • Alzheimer’s Alternative: A new JAMA study detailed a blood test that can detect signs of Alzheimer’s disease before its identifiable with tau-PET imaging. The study of 490 people (225 healthy controls, 89 w/ subjective cognitive decline, 176 w/ mild cognitive impairment) found that the healthy participants who had abnormal Aβ-PET scans and normal tau-PET in their entorhinal cortex also had higher plasma P-tau217 levels.
  • Hologic’s Unifi Update: Hologic announced its updated Unifi Analytics platform (v 1.2) that provides new insights to help users better understand and improve their daily mammography workflow. The updates allow users to quantify study volume and time spent, benchmark their efficiency against other Hologic sites, and measure paddle utilization and compression force.

The Resource Wire

– This is sponsored content.

  • Check out this study detailing Arterys’ Cardio AI solution and how 4D flow is becoming an essential imaging modality for reading cardiovascular disorders.
  • Did you know that one in three Americans is obese and obesity is even more prevalent in rural communities? This Hitachi blog shows how its wide aperture CT and MRI systems are the best fit for rural hospitals, helping them care for patients of all sizes and get more ROI from their imaging systems.
  • ClearRead Xray from Riverain Technologies includes the first FDA-cleared software solution to transform a chest x-ray into a soft-tissue image, providing unprecedented clarity for efficient, accurate, early detection of lung disease.
  • This Healthcare Administrative Partners blog post details how independent radiology practices can build upon their relationships with hospitals, referring physicians, and even neighboring radiology practices to remain strong and independent in the face of ongoing consolidation.
  • This GE Healthcare Insight details why radiology should prioritize cybersecurity, highlighting imaging devices’ emerging risks and central role, and steps to protect them.
  • Learn how one critical access hospital in a California ski mountain town used Nuance PowerShare accelerate care, drive change, and pivot on a dime, and to #ditchthedisk in this upcoming webinar.

Get every issue of The Imaging Wire, delivered right to your inbox.

Join thousands of imaging professionals.