Michigan Medicine’s Ella A. Kazerooni, MD after a Chinese study showed that including low-risk populations in LDCT lung cancer screening programs drives overdiagnosis.

Imaging Wire Sponsors

annalise.ai  •  Arterys  •  Bayer  •  Blackford  •  Canon Medical  •  CARPL.ai  •  Change Healthcare  •  Enlitic  •  GE HealthCare  •  Hyperfine  •  Intelerad  •  Novarad  •  Riverain Technologies  •  Siemens Healthineers  •  Subtle Medical  •  United Imaging  •  Us2.ai

Artificial Intelligence

CXR AI’s Screening Generalizability Gap

A new European Radiology study detailed a commercial CXR AI tool’s challenges when used for screening patients with low disease prevalence, bringing more attention to the mismatch between how some AI tools are trained and how they’re applied in the real world.

The researchers used an unnamed commercial AI tool to detect abnormalities in 3k screening CXRs sourced from two healthcare centers (2.2% w/ clinically significant lesions), and had four radiology residents read the same CXRs with and without AI assistance, finding that the AI:

• Produced a far lower AUROC than in its other studies (0.648 vs. 0.77–0.99)
• Achieved 94.2% specificity, but just 35.3% sensitivity
• Detected 12 of 41 pneumonia, 3 of 5 tuberculosis, and 9 of 22 tumors 
• Only “modestly” improved the residents’ AUROCs (0.571–0.688 vs. 0.534–0.676)
• Added 2.96 to 10.27 seconds to the residents’ average CXR reading times

The researchers attributed the AI tool’s “poorer than expected” performance to differences between the data used in its initial training and validation (high disease prevalence) and the study’s clinical setting (high-volume, low-prevalence, screening).

• More notably, the authors pointed to these results as evidence that many commercial AI products “may not directly translate to real-world practice,” urging providers facing this kind of training mismatch to retrain their AI or change their thresholds, and calling for more rigorous AI testing and trials.

These results also inspired lively online discussions. Some commenters cited the study as proof of the problems caused by training AI with augmented datasets, while others contended that the AI tool’s AUROC still rivaled the residents and its “decent” specificity is promising for screening use.

The Takeaway

We cover plenty of studies about AI generalizability, but most have explored bias due to patient geography and demographics, rather than disease prevalence mismatches. Even if AI vendors and researchers are already aware of this issue, AI users and study authors might not be, placing more emphasis on how vendors position their AI products for different use cases (or how they train it).

Annalise CXR In Action

annalise.ai’s Annalise CXR solution detects up to 124 findings in a single chest X-ray. See how it detects such a wide range of abnormalities using these demo studies…  or upload your own CXR images.

Human Factors in Medical Device Design

How patients interact with a medical device can have as much impact as the device itself. Check out this Q&A with Hyperfine Lead Product Designer Corinne Hay to learn how human factors influence the design of everything from prescription containers to portable MRI systems.

The Wire

• Head & Neck Cancer Surveillance: Huntsman Cancer researchers found that surveillance imaging might reduce two-year mortality rates for certain head and neck cancers. The 1,004 patient-study (902 w/ squamous cell carcinoma) revealed that although surveillance imaging didn’t have a statistically significant impact across the entire sample (HR: 0.76), it was associated with lower mortality among SCC patients with regionalized or distant stage cancers (HRs: 0.55 & 0.40) and among patients with non-SCC cancers (HR: 0.19). 

• Viz.ai Vascular Suite: Viz.ai launched its new Viz Vascular Suite, combining the company’s trademark care coordination capabilities with its existing vascular AI tools (pulmonary embolism, right heart strain, aortic disease detection) and a new FDA-pending abdominal aortic aneurysm solution. The Viz Vascular Suite launches shortly after Viz.ai’s new Radiology Suite and Cardio Suite, as the care coordination AI company increasingly looks to create solution packages for clinical workflows beyond neuro.

• LDCT Overdiagnosis: New research out of China illustrated how including low-risk populations in LDCT screening programs can drive overdiagnosis. A 3M-person screening program (2002-2017, 34k LC cases, 27k deaths) saw early-stage lung cancer diagnosis increase by far more among women than men (+16.1 vs. +6.9 per 100k), but women had much smaller declines in late-stage cancer diagnosis (-0.6 vs. -5.5 per 100k) and no declines in mortality. Noting that 95% of the women were non-smokers, the authors called for changes to screening criteria.

• Echo AI AS Detection: A Tufts Medical Center and iCardio.ai study showed that echo AI could support fully-automated aortic stenosis (AS) detection, and potentially allow AS screening. After training, validation, and testing using 577 patients’ annotated limited 2D echo images, the AI solution accurately performed automated AS screening (AUROC: 0.96), and was able to differentiate “no significant AS” and “early AS” from “significant AS” (AUROCs: 0.86 & 0.75). The echo AI system also accurately screened for AS using 8.5k TTEs from an external dataset (AUROC: 0.91).

• University Hospitals Adopts CARPL.ai: University Hospitals and CARPL.ai launched a multi-year strategic agreement, allowing the prestigious Cleveland area institution to leverage CARPL.ai for radiology AI validation and deployment. The UH Radiology Innovation team will use CARPL’s D.E.V.-D framework (Discovery, Exploration, Validation & Deployment) to quickly access, assess and integrate AI solutions into its RadiCLE ecosystem, and to support its FDA clearance validation services.

• EHR “Better Practice” Nudges: A JAMIA study showed how changes to EHR “choice architecture” can reduce incorrect imaging orders. After finding that providers were incorrectly ordering “CT abdomen” exams instead of “CT abdomen/pelvis” exams, they adjusted their imaging order naming structure so providers saw “CT abdomen /pelvis” first. Similarly, after noticing that patients were being over-sedated for imaging exams because the EHR defaulted to doses intended for patients who regularly take anxiety medications, they added a new order specifically for imaging patients.

• December Jobs Report: The December Jobs Report revealed that the healthcare sector added 55k jobs last month, with ambulatory services, hospitals, and residential care facilities leading the job gains (30k,16k, 9k). The healthcare industry’s been surprisingly quick to make up the job losses from the beginning of the pandemic, adding an average of 49k jobs per month in 2022 (vs. 9k per month in 2021), although it’s unclear if the momentum will continue now that the COVID job recovery is in the rearview mirror.

• Early Infarct AI: An MGH-led team developed a DL model that accurately detected early acute infarcts in non-contrast head CTs, which is typically challenging with CT and diagnosed using MRI. The AI outperformed three neurorads with 150 CTs (sensitivity: 96% vs. 61-66%; specificity: 72% vs. 90-92%), and when the model was applied to an expanded 364 CT dataset it detected >70 mL infarcts with 97% sensitivity and 99% specificity. 

• Onc.AI Scores $25M: Onc.AI raised $25M in Series A funding to develop and obtain regulatory approval for its upcoming precision-oncology platform, which leverages medical data and imaging radiomics to support cancer treatment decision-making. Onc.AI is part of a small but growing group of multimodal AI startups targeting clinical and life science use cases.

• Brain Aneurysm AI’s Research Problem: King’s College London researchers added to the growing list of critical AI study reviews, finding widespread challenges across 43 previous brain aneurysm AI studies. The researchers identified high risks of bias and poor generalizability across most studies (26% used ideal reference standards, 14% tested externally) and low true-positive rates (89%). Although some of the studies were from AI’s early days and it’s unclear how many were commercial products, the authors concluded that the overall aneurysm AI segment isn’t ready for clinical use.

• Samsung S-Detects’ Thyroid Performance: A new European Radiology study highlighted Samsung’s S-Detect thyroid cancer ultrasound AI solution’s solid performance. Researchers used S-Detect with 236 patients (w/ 312 suspicious nodules), achieving sensitivity, accuracy, and AUCs (0.95, 0.84, & 0.753) that were comparable to three senior radiologists and higher than nine residents. When used as an assistant, S-Detect also improved the residents’ detection of ≤1.5 cm nodules (all p < 0.01) and decreased unnecessary biopsies of >1.5 cm nodules by up to 27.7%.

• Gradient Adds Cone Health: Gradient Health launched a collaboration with Cone Health, allowing the North Carolina health system to securely/anonymously transfer its “previously unused” healthcare data to AI and life science innovators, while ensuring that its patient population is represented in future developments.

Canon’s Value-Based PET/CT

Canon’s Cartesion Prime is the only air-cooled digital PET/CT system, which means big reductions in service, chiller, and infrastructure costs. Check out Canon’s case for the air-cooled Cartesion Prime.

How Data Fluent is Your Organization?

Healthcare’s data fluency challenges have existed for years, and they are increasingly getting in the way of care delivery and the completion of AI projects. This Enlitic report details the three data fluency challenges that healthcare must overcome, and how it addresses these challenges.

The Resource Wire

• Efficiency and quality are the name of the game at RadNet, and that’s exactly what the imaging center giant achieved when it adopted Subtle Medical’s SubtleMR solution, optimizing its already-accelerated MRI protocols by 33-45% while maintaining consistent diagnostic image quality.

• Change Healthcare’s cloud-native, zero-footprint Stratus Imaging PACS is live in clinical use. See how Stratus Imaging PACS is helping radiology practices improve productivity and patient care, while eliminating the cost and resource constraints of on-premise systems.

• See how Einstein Healthcare Network reduced its syringe expenses, enhanced its syringe loading, and improved its contrast documentation when it upgraded to Bayer Radiology’s MEDRAD Stellant FLEX CT Injection System.

• Learn why Texas’ Baylor Scott & White Medical Center and the UK’s Queen Elizabeth Hospital Birmingham decided to upgrade their SPECT-only cameras and first-generation SPECT/CTs to Siemens Healthineers’ Symbia Pro.specta SPECT/CT and how they’ve benefited since then.

• Think your patients are ready for you to ditch the disk? Join ACR past president, Dr. Geraldine McGinty, and Intelerad president, Morris Panner, for this Imaging Wire Show exploring the state of image sharing, its impact on patient care, and how radiology can finally ditch the disk.

• United Imaging’s “all-in” approach means that every system ships with its entire suite of features and capabilities (no options), giving its clients more clinical flexibility and predictability.

• Have more echo studies than sonographers? See how Us2.ai was able to classify, segment, and annotate echocardiographic videos with similar accuracy as expert sonographers. 

• Imaging AI is evolving fast, but radiology leaders’ expectations for their AI technologies might be evolving even faster. In this Imaging Wire Show with Dr. Charlene Liew of SingHealth and Dr. Nina Kottler of Radiology Partners, we explore radiology leaders’ current and future expectations for AI, and the central role platforms play in their AI roadmaps.

• When this 66 year-old woman was referred for pain and functional impotence of the wrist, her initial X-ray images were normal. However, Arterys’ Chest I MSK AI detected a fracture on the dorsal side of her cortical bone, alerting the radiologist and confirming her injury.

• Imaging providers who want to finally #ditchthedisk can now start off with Novarad’s CryptoChart Lite solution, a standard version of CryptoChart built for providers transitioning to imaging sharing.

• We may be on the verge of the next revolution in how cancer is detected, diagnosed, and treated and this editorial by GE HealthCare Imaging leader, Jan Makela, details how GE and its partners are using AI and imaging data to drive this revolution.

• This Riverain Technologies case study details how Einstein Medical Center adopted ClearRead CT enterprise-wide (all 13 CT scanners) and how the solution allowed Einstein radiologists to identify small nodules faster and more reliably.

• Working out your AI business case? Check out this helpful Blackford Analysis post outlining the three steps to help you justify AI adoption.