The Case for Algorithmic Audits

A new Lancet Digital Health study could have become one of the many “AI rivals radiologists” papers that we see each week, but it instead served as an important lesson that traditional performance tests might not prove that AI models are actually safe for clinical use.

The Model – The team developed their proximal femoral fracture detection DL model using 45.7k frontal X-rays performed at Australia’s Royal Adelaide Hospital (w/ 4,861 fractures).

The Validation – They then tested it against a 4,577-exam internal set (w/ 640 fractures), 400 of which were also interpreted by five radiologists (w/ 200 fractures), and against an 81-image external validation set from Stanford.

The Results – All three tests produced results that a typical study might have viewed as evidence of high-performance: 

  • The model outperformed the five radiologists (0.994 vs. 0.969 AUCs)
  • It beat the best performing radiologist’s sensitivity (95.5% vs. 94.5%) and specificity (99.5% vs 97.5%)
  • It generalized well with the external Stanford data (0.980 AUC)

The Audit – Despite the strong results, a follow-up audit revealed that the model might make some predictions for the wrong reasons, suggesting that it is unsafe for clinical deployment:

  • One false negative X-ray included an extremely displaced fracture that human radiologists would catch
  • X-rays featuring abnormal bones or joints had a 50% false negative rate, far higher than the reader set’s overall false negative rate (2.5%)
  • Salience maps showed that AI decisions were almost never based on the outer region of the femoral neck, even with images where that region was clinically relevant (but it still often made the right diagnosis)
  • The model scored a high AUC with the Stanford data, but showed a substantial model operating point shift

The Case for Auditing – Although the study might have not started with this goal, it ended up becoming an argument for more sophisticated preclinical auditing. It even led to a separate paper outlining their algorithmic auditing process, which among other things suggested that AI users and developers should co-own audits.

The Takeaway

Auditing generally isn’t the most exciting topic in any field, but this study shows that it’s exceptionally important for imaging AI. It also suggests that audits might be necessary for achieving the most exciting parts of AI, like improving outcomes and efficiency, earning clinician trust, and increasing adoption.A new Lancet Digital Health study could have become one of the many “AI rivals radiologists” papers that we see each week, but it instead served as an important lesson that traditional performance tests might not prove that AI models are actually safe for clinical use.

Imaging AI’s Big 2021

Signify Research’s latest imaging AI VC funding report revealed an unexpected surge in 2021, along with major funding shifts that might explain why many of us didn’t see it coming. Here’s some of Signify’s big takeaways and here’s where to get the full report.

AI’s Path to $3.47B – Imaging AI startups have raised $3.47B in venture funding since 2015, helped by a record-high $815M in 2021 after several years of falling investments (vs. 2020’s $592M, 2019’s $450M, 2018’s $790M).

Big Get Bigger – That $3.47B funding total came from over 200 companies and 290 deals, although the 25 highest-funded companies were responsible for 80% of all capital raised. VCs  increased their focus on established AI companies in 2021, resulting in record-high late-stage funding (~$723.5M), record-low Pre-Seed/Seed funding (~$7M), and a major increase in average deal size (~$33M vs. ~$12M in 2020). 

Made in China – If you’re surprised that 2021 was a record AI funding year, that’s probably because it targeted Chinese companies (~$260M vs. US’ ~$150M), continuing a recent trend (China’s AI VC share was 45% in 2020, 26% in 2019). We’re also seeing major funding go to South Korea and Australia’s top startups, adding to APAC AI vendors’ funding leadership.

Health VC Context – Although imaging AI’s $815M 2021 funding total seems big for a category that’s figuring out its path towards full adoption, the amount VC firms are investing in other areas of healthcare makes it seem pretty reasonable. Our two previous Digital Health Wire issues featured seven digital health startup funding rounds with a total value of $267M (and that’s from just one week).

The Takeaway

Signify correctly points out that imaging AI funding remains strong despite a list of headwinds (COVID, regulatory hurdles, lacking reimbursements), while showing more signs of AI market maturation (larger funding rounds to fewer players) and suggesting that consolidation is on the way. Those factors will likely continue in 2022. However, more innovation is surely on the way too and quite a few regional AI powerhouses still haven’t expanded globally, suggesting that the next steps in AI’s evolution won’t be as straightforward as some might think.

Autonomous AI Milestone

Just as the debate over whether AI might replace radiologists is starting to fade away, Oxipit’s ChestLink solution became the first regulatory-approved imaging AI product intended to perform diagnoses without involving radiologists (*please see editor’s note below regarding Behold.ai). That’s a big and potentially controversial milestone in the evolution of imaging AI and it’s worth a deeper look.

About ChestLink – ChestLink autonomously identifies CXRs without abnormalities and produces final reports for each of these “normal” exams, automating 15% to 40% of reporting workflows.

Automation Evidence – Oxipit has already piloted ChestLink in supervised settings for over a year, processing over 500k real-world CXRs with 99% sensitivity and no clinically relevant errors.

The Rollout – With its CE Class IIb Mark finalized, Oxipit is now planning to roll out ChestLink across Europe and begin “fully autonomous” operation by early 2023. Oxipit specifically mentioned primary care settings (many normal CXRs) and large-scale screening projects (high volumes, many normal scans) in its announcement, but ChestLink doesn’t appear limited to those use cases.

ChestLink’s ability to address radiologist shortages and reduce labor costs seem like strong and unique advantages. However, radiology’s first regulatory approved autonomous AI solution might face even stronger challenges:

  • ChestLink’s CE Mark doesn’t account for country-specific regulations around autonomous diagnostic reporting (e.g. the UK requires “appropriate reporting” with ionizing radiation-based exams)
  • Radiologist societies historically push back against anything that might undermine radiologists’ clinical roles, earning potential, and future career stability
  • Health systems’ evidence requirements for any autonomous AI tools would likely be extremely high, and they might expect similarly high economic ROI in order to justify the associated diagnostic or reputational risks
  • Even the comments in Oxipit’s LinkedIn announcement had a much more skeptical tone than we typically see with regulatory approval announcements

The Takeaway

Autonomous AI products like ChestLink could address some of radiology’s greatest problems (radiologist overwork, staffing shortages, volume growth, low access in developing countries) and their economic value proposition is far stronger than most other diagnostic AI products.

However, autonomous AI solutions could also face more obstacles than any other imaging AI products we’ve seen so far, suggesting that it would take a combination of excellent clinical performance and major changes in healthcare policies/philosophies in order for autonomous AI to reach mainstream adoption.

*Editor’s Note – April 21, 2022: Behold.ai insists that it is the first imaging AI company to receive regulatory approval for autonomous AI. Its product is used with radiologist involvement and local UK guidelines require that radiologists read exams that use ionizing radiation. All above analysis regarding the possibilities and challenges of autonomous AI applies to any autonomous AI vendor in the current AI environment, including both Oxipit and Behold.ai.

Complementary PE AI

A new European Radiology study out of France highlighted how Aidoc’s pulmonary embolism AI solution can serve as a valuable emergency radiology safety net, catching PE cases that otherwise might have been missed and increasing radiologists’ confidence. 

Even if that’s technically what PE AI products are supposed to do, studies using commercially available products and focusing on how AI complements radiologists (vs. comparing AI and rad accuracy) are still rare and worth a closer look.

The Diagnostic Study – A team from French telerad provider, IMADIS, analyzed AI and radiologist CTPA interpretations from patients with suspected PE (n = 1,202 patients), finding that:

  • Aidoc PE achieved higher sensitivity (0.926 vs. 0.9 AUCs) and negative predictive value (0.986 vs. 0.981 AUCs)
  • Radiologists achieved higher specificity (0.991 vs. 0.958 AUCs), positive predictive value (0.95 vs. 0.804 AUCs), and accuracy (0.977 vs. 0.953 AUCs)
  • The AI tool flagged 219 suspicious PEs, with 176 true positives, including 19 cases that were missed by radiologists
  • The radiologists detected 180 suspicious PEs, with 171 true positives, including 14 cases that were missed by AI
  • Aidoc PE would have helped IMADIS catch 285 misdiagnosed PE cases in 2020 based on the above AI-only PE detection ratio (19 per 1,202 patients)  

The Radiologist Survey – Nine months after IMADIS implemented Aidoc PE, a survey of its radiologists (n = 79) and a comparison versus its pre-implementation PE CTPAs revealed that:

  • 72% of radiologists believed Aidoc PE improved their diagnostic confidence and comfort 
  • 52% of radiologists the said the AI solution didn’t impact their interpretation times
  • 14% indicated that Aidoc PE reduced interpretation times
  • 34% of radiologists believed the AI tool added time to their workflow
  • The solution actually increased interpretation times by an average of 7.2% (+1:03 minutes) 

The Takeaway

Now that we’re getting better at not obsessing over AI replacing humans, this is a solid example of how AI can complement radiologists by helping them catch more PE cases and make more confident diagnoses. Some radiologists might be concerned with false positives and added interpretation times, but the authors noted that AI’s PE detection advantages (and the risks of missed PEs) outweigh these potential tradeoffs.

Sirona Medical Acquires Nines AI, Talent

Sirona Medical announced its acquisition of Nines’ AI assets and personnel, representing notable milestones for Sirona’s integrated RadOS platform and the quickly-changing imaging AI landscape.

Acquisition Details – Sirona acquired Nines’ AI portfolio (data pipeline, ML engines, workflow/analytics tools, AI models) and key team members (CRO, Direct of Product, AI engineers), while Nines’ teleradiology practice was reportedly absorbed by one of its telerad customers. Terms of the acquisition weren’t disclosed, although this wasn’t a traditional acquisition considering that Sirona and Nines had the same VC investor.

Sirona’s Nines Strategy – Sirona’s mission is to streamline radiologists’ overly-siloed workflows with its RadOS radiology operating system (unifies: worklist, viewer, reporting, AI, etc.), and it’s a safe bet that any acquisition or investment Sirona makes is intended to advance this mission. With that…

  • Nine’s most tangible contributions to Sirona’s strategy are its FDA-cleared AI models: NinesMeasure (chest CT-based lung nodule measurements) and NinesAI Emergent Triage (head CT-based intracranial hemorrhage and mass effect triage). The AI models will be integrated into the RadOS platform, bolstering Sirona’s strategy to allow truly-integrated AI workflows. 
  • Nine’s personnel might have the most immediate impact at Sirona, given the value/scarcity of experienced imaging software engineers and the fact that Nines’ product team arguably has more hands-on experience with radiologist workflows than any other imaging AI firm (at least AI firms available for acquisition).
  • Nine’s other AI and imaging workflow assets should also help support Sirona’s future RadOS and AI development, although it’s harder to assess their impact for now.

The AI Shakeup Angle – This acquisition has largely been covered as another example of 2022’s AI shakeup, which isn’t too surprising given how active this year has been (MaxQ’s shutdown, RadNet’s Aidence/Quantib acquisitions, IBM shedding Watson Health). However, Nines’ strategy to combine a telerad practice with in-house AI development was quite unique and its decision to sell might say more about its specific business model (at its scale) than it does about the overall AI market.

The Takeaway

Since the day Sirona emerged from stealth, it’s done a masterful job articulating its mission to solve radiology’s workflow problems by unifying its IT infrastructure. Acquiring Nines’ AI assets certainly supports Sirona’s unified platform messaging, while giving it more technology and personnel resources to try to turn that message into a reality.

Meanwhile, Nines becomes the latest of surely many imaging AI startups to be acquired, pivoted, or shut down, as AI adoption evolves at a slower pace than some VC runways. Nines’ strategy was really interesting, they had some big-name founders and advisors, and now their work and teams will live on through Sirona.

Intracranial Hemorrhage AI Efficiency

A new Radiology: Artificial Intelligence study out of Switzerland highlighted how Aidoc’s Intracranial Hemorrhage AI solution improved emergency department workflows, without hurting patient care. Even if that’s exactly what solutions like this are supposed to do, real world AI studies that go beyond sensitivity and specificity are still rare and worth some extra attention.

The Study – The researchers analyzed University Hospital of Basel’s non-contrast CT intracranial hemorrhage (ICH) exams before and after adopting the Aidoc ICH solution (n = 1,433 before & 3,017 after; ~14% ICH incidence w/ both groups).

Diagnostic Results – The Aidoc solution produced “practicable” overall diagnostic results (93% accuracy, 87.2% sensitivity, 93.9% specificity, and 97.8% NPV), although accuracy was lower with certain ICH subtypes (e.g. subdural hemorrhage 69.2%, 74/107). 

Efficiency Results – More notably, the Aidoc ICH solution “positively impacted” UBS’ ED workflows, with improvements across a range of key metrics:

  • Communicating critical findings: 63 vs. 70 minutes
  • Communicating acute ICH: 58 vs. 73 minutes
  • Overall turnaround time to rule out ICH: 164 vs. 175 minutes
  • Turnaround time to rule out ICH during working hours: 167 vs. 205 minutes

Next Steps – The authors called for further efforts to streamline their stroke workflows and to create a clear ICH AI framework, accurately noting that “AI tools are only as reliable as the environment they are deployed in.”

The Takeaway
The internet hasn’t always been kind to emergency AI tools, and academic studies have rarely focused on the workflow efficiency outcomes that many radiologists and emergency teams care about. That’s not the case with this study, which did a good job showing the diagnostic and workflow upsides of ICH AI adoption, and added a nice reminder that imaging teams share responsibility for AI outcomes.

Creating a Cancer Screening Giant

A few days after shocking the AI and imaging center industries with its acquisitions of Aidence and Quantib, RadNet’s Friday investor briefing revealed a far more ambitious AI-enabled cancer screening strategy than many might have imagined.

Expanding to Colon Cancer – RadNet will complete its AI screening platform by developing a homegrown colon cancer detection system, estimating that its four AI-based cancer detection solutions (breast, prostate, lung, colon) could screen for 70% of cancers that are imaging-detectable at early stages.

Population Detection – Once its AI platform is complete, RadNet plans to launch a strategy to expand cancer screening’s role in population health, while making prostate, lung, and colon cancer screening as mainstream as breast cancer screening.

Becoming an AI Vendor – RadNet revealed plans to launch an externally-focused AI business that will lead with its multi-cancer AI screening platform, but will also create opportunities for RadNet’s eRAD PACS/RIS software. There are plenty of players in the AI-based cancer detection arena, but RadNet’s unique multi-cancer platform, significant funding, and training data advantage would make it a formidable competitor.

Geographic Expansion – RadNet will leverage Aidence and Quantib’s European presence to expand its software business internationally, as well as into parts of the US where RadNet doesn’t own imaging centers (RadNet has centers in just 7 states).

Imaging Center Upsides – RadNet’s cancer screening AI strategy will of course benefit its core imaging center business. In addition to improving operational efficiency and driving more cancer screening volumes, RadNet believes that the unique benefits of its AI platform will drive more hospital system joint ventures.

AI Financials – The briefing also provided rare insights into AI vendor finances, revealing that DeepHealth has been running at a $4M-$5M annual loss and adding Aidence / Quantib might expand that loss to $10M- $12M (seems OK given RadNet’s $215M EBITDA). RadNet hopes its AI division will become cash flow neutral within the next few years as revenue from outside companies ramp up.

The Takeaway

RadNet has very big ambitions to become a global cancer screening leader and significantly expand cancer screening’s role in society. Changing society doesn’t come fast or easy, but a goal like that reveals how much emphasis RadNet is going to place on developing and distributing its AI cancer screening platform going forward.

Duke’s Interpretable AI Milestone

A team of Duke University radiologists and computer engineers unveiled a new mammography AI platform that could be an important step towards developing truly interpretable AI.

Explainable History – Healthcare leaders have been calling for explainable imaging AI for some time, but explainability efforts have been mainly limited to saliency / heat maps that show what part of an image influenced a model’s prediction (not how or why).

Duke’s Interpretable Model – Duke’s new AI platform analyzes mammography exams for potentially cancerous lesions to help physicians determine if a patient should receive a biopsy, while supporting its predictions with image and case-based explanations. 

Training Interpretability – The Duke team trained their AI platform to locate and evaluate lesions following a process that human radiology educators and students would utilize:

  • First, they trained the AI model to detect suspicious lesions and to ignore healthy tissues
  • Then they had radiologists label the edges of the lesions
  • Then they trained the model to compare those lesion edges with lesion edges from an archive of images with confirmed outcomes

Interpretable Predictions – This training process allowed the AI model to identify suspicious lesions, highlight the classification-relevant parts of the image, and explain its findings by referencing confirmed images. 

Interpretable Results – Like many AI models, this early version could not identify cancerous lesions as accurately as human radiologists. However, it matched the performance of existing “black box” AI systems and the team was able to see why their AI model made its mistakes.

The Takeaway

It seems like concerns over AI performance are growing at about the same pace as actual AI adoption, making explainability / interpretability increasingly important. Duke’s interpretable AI platform might be in its early stages, but its use of previous cases to explain findings seems like a promising (and straightforward) way to achieve that goal, while improving diagnosis in the process.

The False Hope of Explainable AI

Many folks view explainability as a crucial next step for AI, but a new Lancet paper from a team of AI heavyweights argues that explainability might do more harm than good in the short-term, and AI stakeholders would be better off increasing their focus on validation.

The Old Theory – For as long as we’ve been covering AI, really smart and well-intentioned people have warned about the “black-box” nature of AI decision making and forecasted that explainable AI will lead to more trust, less bias, and greater adoption.

The New Theory – These black-box concerns and explainable AI forecasts might be logical, but they aren’t currently realistic, especially for patient-level decision support. Here’s why:

  • Explainability methods describe how AI systems work, not how decisions are made
  • AI explanations can be unreliable and/or superficial
  • Most medical AI decisions are too complex to explain in an understandable way
  • Humans over-trust computers, so explanations can hurt their ability to catch AI mistakes
  • AI explainability methods (e.g heat maps) require human interpretation, risking confirmation bias
  • Explainable AI adds more potential error sources (AI tool + AI explanation + human interpretation)
  • Although we still can’t fully explain how acetaminophen works, we don’t question whether it works, because we’ve tested it extensively

The Explainability Alternative – Until suitable explainability methods emerge, the authors call for “rigorous internal and external validation of AI models” to make sure AI tools are consistently making the right recommendations. They also advised clinicians to remain cautious when referencing AI explanations and warned that policymakers should resist making explainability a requirement. 

Explability’s Short-Term Role – Explainability definitely still has a role in AI safety, as it’s “incredibly useful” for model troubleshooting and systems audits, which can improve model performance and identify failure modes or biases.

The Takeaway – It appears we might not be close enough to explainable AI to make it a part of short-term AI strategies, policies, or procedures. That might be hard to accept for the many people who view the need for AI explainability as undebatable, and it makes AI validation and testing more important than ever.

Who Owns AI Evaluation and Monitoring?

Imaging AI evaluation and monitoring just became even hotter topics, following a particularly revealing Twitter thread and a pair of interesting new papers.

Rads Don’t Work for AI – A Mayo Clinic Florida neuroradiologist took his case to Twitter after an FDA-approved AI tool missed 6 of 7 hemorrhages in a single shift and he was asked to make extra clicks to help improve the algorithm. No AI tool is perfect, but many folks commenting on this thread didn’t take kindly to the idea of being asked to do pro-bono work to improve an algorithm that they already paid for. 

AI Takes Work – A few radiologists with strong AI backgrounds clarified that this “extra work” is intended to inform the FDA about postmarket performance, while monitoring healthcare tools and providing feedback is indeed physicians’ job. They also argued that radiology practices should ensure that they have the bandwidth to monitor AI before deciding to adopt it.

The ACR DSI Gets It – Understanding that “AI algorithms may not work as expected when used beyond the institutions in which they were trained, and model performance may degrade over time” the ACR Data Science Institute (DSI) released a helpful paper detailing how radiologists can evaluate AI before and during clinical use. In an unplanned nod to the above Twitter thread, the DSA paper also noted that AI evaluation/monitoring is “ultimately up to the end users” although many “practices will not be able to do this on their own.” The good news is the ACR DSI is developing tools to help them.

DLIR Needs Evaluation Too – Because measuring whether DL-reconstructed scans “look good” or allow reduced dosage exams won’t avoid errors (e.g. false tumors or removed tumors), a Washington University in St. Louis-led team is developing a framework for evaluating DLIR tools before they are introduced into clinical practice. The new framework comes from some big-name intuitions (WUSTL, NIH, FDA, Cleveland Clinic, UBC), all of whom also appear to agree that AI evaluation is up to the users.

The Takeaway – At least among AI insiders it’s clear that AI users are responsible for algorithm evaluation and monitoring, even if bandwidth is limited and many evaluation/monitoring tools are still being developed. Meanwhile, many AI users (who are crucial for AI to become mainstream) want their FDA-approved algorithms to perform correctly and aren’t eager to do extra work to help improve them. That’s a pretty solid conflict, but it’s also a silver lining for AI vendors who get good at streamlining evaluations and develop low-labor ways to monitor performance.

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

You're signed up!

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

-- The Imaging Wire team