Anthropic Calls For Pause Or Slowdown In AI Development After Pope Leo’s Encyclical
By EWTN News
By Tyler Arnold
Less than two weeks after Pope Leo XIV published an encyclical warning artificial intelligence (AI) companies against constructing “a new Tower of Babel,” the multibillion-dollar AI company Anthropic is calling for a global pause or slowdown in development.
Anthropic co-founder Jack Clark and Anthropic Institute head Marina Favaro published a blog on June 4 warning about a risk of “humans losing control over AI systems” as its own system Claude is reaching the potential to autonomously design its own successor without any human contributions.
“This is called recursive self-improvement,” they wrote. “We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for.”
The blog post did not mention the encyclical, but a separate Anthropic co-founder, Chris Olah, met with Leo and sat alongside the pope when the encyclical Magnifica Humanitas was revealed on May 25. Anthropic has engaged in outreach to the Vatican and other religious leaders to help address ethical questions related to AI development.
In the blog post, Anthropic leaders explained that its AI system is taking over a large portion of writing code that designs AI — with its workload growing eightfold every quarter. AI will “become much more capable in coming years,” they wrote, and “these trends have huge implications.”
“If systems are capable of fully building their own successors, the ways we secure them, monitor them, and shape their behavior all grow much more important,” they wrote.
Although Clark and Favaro acknowledged AI has not reached this level yet and they cannot say for certain it will, they wrote: “We do not have good intuitions for what this world would look like” if this occurs, and AI capabilities “eclipse those of humans.”
Anthropic’s leaders wrote that AI companies should come together to either pause or slow down development “to give ourselves more time to deal with its immense implications.” However, this would require global international cooperation among countries and AI companies because “if a slowdown simply lets the least cautious actors catch up technologically, it could leave everyone less safe,” they wrote.
“We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology,” they added.
Anthropic intends to engage with policymakers, researchers, and other members of the public to discuss these concerns. The company will publish a document based on what comes out of the conversations.
‘Disarming’ AI
Charles Camosy, a moral theologian at The Catholic University of America who has worked with Anthropic on ethical questions, told EWTN News that Anthropic’s statements appear in line with Leo’s desire to “disarm” AI, which the pontiff explained as not halting innovation but “preventing it from dominating humanity.”
He said Anthropic recognizes the speed of development as “such a problem we all need to slow down here.” Such a pause would allow society to “think about what AI should or should not do in the culture,” he said.
Camosy pointed to concerns about “outsourcing” teaching, tutoring, parenting, care for the sick, and other human interactions to AI, possibly “undermining the things that … make our humanity magnificent.”
He recognized that fierce AI competition among nations and companies “creates a significant roadblock” to global cooperation for slowing everything down, but said: “I’ve been astonished by how many different kinds of people are interested in this encyclical.”
“Many people were kind of waiting for someone to fill the moral space,” Camosy said and suggested the Church help lead a global movement that demands ethical AI, and he encouraged the Holy Father to consider a trip to Silicon Valley.
“To many people that sounds hopefully naive,” he said. “But I don’t see another choice here.”
Mayo Clinic study shows AI can reveal brain tumor risks without costly genetic testing
ROCHESTER, Minn. — Mayo Clinic researchers and collaborators have shown that artificial intelligence (AI) can analyze routine pathology slides to help classify meningiomas, the most common primary brain tumor in adults, and predict a patient's risk of tumor recurrence.
The study, published in The Lancet Digital Health, demonstrates that deep learning models can extract molecular and prognostic information from standard hematoxylin and eosin, or H&E, slides — the same type of tissue images already used in routine clinical care. These insights are typically obtained through DNA methylation profiling, an advanced genetic test which provides valuable diagnostic and prognostic information but can be costly, time-consuming and is unavailable in many hospitals.
"This is one of the many studies where we can harness the strength of digital pathology by capturing the last two decades of genomic and molecular knowledge into AI algorithms," says Gelareh Zadeh, M.D., Ph.D., chair of the Department of Neurologic Surgery at Mayo Clinic in Rochester and the David C. and Flora C. Pratt Distinguished Chief Medical Officer for Mayo Clinic Platform.
Making advanced tumor insights more accessible
Meningiomas can vary widely in behavior. Some grow slowly and may never return after treatment, while others are more aggressive and more likely to recur. Understanding that risk is critical for patients and care teams deciding whether additional treatment, such as radiation therapy, may be needed after surgery.
Molecular testing can help identify which tumors are more likely to recur and which may respond differently to treatment. But these tests require specialized technology and expertise, limiting access for many patients.
Using tissue samples, pathology images and clinical data from 672 patients, researchers trained AI to uncover information about a tumor's biology. Drawing on multiple de-identified datasets, including data resources from Mayo Clinic Platform, the models were able to classify meningioma subtypes and predict recurrence risk using standard pathology slides that are already part of routine patient care.
The findings suggest that AI could one day help clinicians obtain more detailed tumor information without requiring patients to undergo advanced genetic testing.
Helping guide treatment decisions
For patients with meningiomas, recurrence risk can influence follow-up care, imaging frequency and whether radiation therapy should be considered. The study found that AI-based predictions remained useful even after accounting for traditional clinical factors such as tumor grade, the extent to which surgery was able to remove the tumor and patient age.
Researchers also found that the AI models could identify patterns of tumor heterogeneity — differences within the same tumor — that may help explain why some tumors behave more aggressively or respond differently to treatment.
The researchers note that additional prospective studies are needed before the AI models can be used routinely in clinical care. Still, they say the findings lay the groundwork for more accessible, personalized care for patients with meningiomas — and potentially for similar AI approaches in other cancers.
"The aim is to make these algorithms readily and simply accessible for use globally, improving patient care across many healthcare settings," says Dr. Zadeh.
For a complete list of authors, disclosures and funding, review the publication.
###
About Mayo Clinic
Mayo Clinic is a nonprofit organization committed to innovation in clinical practice, education and research, and providing compassion, expertise and answers to everyone who needs healing. Visit the Mayo Clinic News Network for additional Mayo Clinic news.
About Mayo Clinic Platform
Founded on Mayo Clinic's dedication to patient-centered care, Mayo Clinic Platform enables new knowledge, new solutions, and new technologies through collaborations with health technology innovators to create a healthier world. To learn more, visit Mayo Clinic Platform at www.mayoclinicplatform.org.
Journal
The Lancet Digital Health
Article Publication Date
5-Jun-2026
What AI hiring systems reveal about the future of work
Jon Stojan
June 3, 2026
Opinions expressed by Digital Journal contributors are their own.
AI is quietly rewriting the rules of hiring. Not by replacing recruiters entirely, but by changing the very first conversation candidates have with a company.
Imagine applying for a job and skipping the resume black hole altogether. No keyword stuffing. No guessing whether your college, accent, or background will unconsciously sway a recruiter. Instead, your first interaction is a live conversation with an AI interviewer trained to evaluate your skills, problem-solving ability, and communication in real time.
That future is already here.
Companies like HackerEarth are helping some of the world’s largest tech employers rethink how talent is identified and assessed. Their AI-powered screening and interview tools are designed to move beyond resumes and surface candidates based on demonstrated ability rather than pedigree alone. In an industry where employers often face challenges filling technical roles efficiently, qualified candidates are sometimes overlooked.
But the rise of AI interviews also forces a deeper conversation about what meritocracy really means in a digital economy.
For decades, hiring has operated on imperfect signals. Brand-name universities, polished resumes, referrals, and even geography have played outsized roles in determining who gets noticed. The problem is that these signals don’t always correlate with talent. They correlate with access.
That gap has become impossible to ignore in a world where brilliant engineers can emerge from Lagos, São Paulo, or a rural town in India just as easily as Silicon Valley. Remote work cracked open the global talent market. AI may be the force that finally restructures it.
In theory, AI-driven interviews could create a more level playing field. Supporters of these systems argue that, when carefully designed and monitored, AI tools can place greater emphasis on candidates’ skills, communication, and problem-solving abilities rather than traditional signals such as educational background, location, or professional networks. However, outcomes can vary depending on the data, design, and oversight behind the system. A well-designed system can focus on how candidates think, communicate, and solve problems under pressure.
That matters because traditional hiring often rewards polish over potential.
Anyone who has spent time in the tech world knows some exceptional builders never learned how to “interview well” in the conventional sense. At the same time, some candidates know how to game the process without necessarily being the best fit for the role. AI screening tools are often intended to reduce that disconnect by introducing more standardized evaluations at scale, although their effectiveness depends on how they are implemented and monitored.
Still, fairness in hiring is not as simple as swapping humans for algorithms.
AI systems are only as unbiased as the data and assumptions behind them. If companies train models on historical hiring patterns, they risk reinforcing the same inequities they claim to eliminate. Bias doesn’t disappear when it becomes automated. In some cases, it becomes harder to detect.
That tension is why the conversation around AI hiring is becoming so important. The real opportunity isn’t replacing human judgment. For many organizations, the goal is not to remove people from the process but to use AI as a decision-support tool while maintaining human oversight and review. It’s building systems where AI handles consistency and scale while humans provide context, empathy, and nuance.
There are also broader questions around transparency, candidate consent, and accountability. Candidates may not always understand how AI-assisted evaluations are conducted, what data is being collected, or how assessment results influence hiring decisions. Employers using these systems may need to establish clear disclosure practices, regularly audit outcomes for unintended bias, and validate AI-generated assessments against real-world hiring and job performance outcomes rather than relying on automated recommendations alone.
The companies getting this right understand that hiring is ultimately about people, not just productivity metrics.
There’s also a psychological shift happening among candidates themselves. Younger workers entering the labor force are increasingly comfortable interacting with AI systems in everyday life. They use AI to write, research, learn, and communicate. For many of them, talking to an AI interviewer may feel less intimidating than speaking to a panel of strangers.
That comfort level could fundamentally reshape candidate behavior. People may prepare differently, communicate differently, and even rethink how they present their careers. Instead of optimizing resumes for recruiters, they may optimize for dynamic conversations and demonstrable skills.
AI-driven evaluation systems are beginning to influence education, professional certification, and internal promotions. As organizations increasingly rely on machine-assisted decision-making, society will need to decide which human qualities matter most in an automated economy.
Can curiosity be measured?
Can resilience?
Can leadership potential emerge through data patterns alone?
These are not abstract questions anymore. They are operational decisions companies are making right now.
The irony is that AI may end up forcing businesses to become more human in the long run. When machines take over repetitive evaluation tasks, companies will need to place greater emphasis on qualities that algorithms struggle to fully quantify: creativity, emotional intelligence, adaptability, and ethical judgment.
That may ultimately be the most important shift of all.
The future of hiring isn’t about humans versus AI. It’s about whether technology can help uncover talent that the old system routinely ignored. If implemented thoughtfully, AI-assisted hiring tools may help organizations evaluate candidates using a broader range of skills and assessments, potentially expanding access to opportunities for some applicants. Achieving those outcomes, however, depends on ongoing monitoring, transparency, validation, and safeguards designed to identify and address unintended bias.
If done carelessly, it could institutionalize a new generation of invisible barriers.
Either way, the interview room is changing. And increasingly, the first voice candidates hear may not be human at all.
Hollywood studios and actors’ union find common ground on AI
AFP
May 31, 2026
As Hollywood’s performers cast their ballots to approve the latest negotiated contract, union leaders say they have made some progress in conversations with studio bosses since the massive strike in 2023, especially when it comes to concerns about artificial intelligence.
SAG-AFTRA chief negotiator Duncan Crabtree-Ireland attributed the mostly drama-free agreements in this round of negotiations to a new mindset, “because the studios and streamers came to the table with a different perspective.”
With 160,000 members working in film, television and video games, SAG-AFTRA is the largest and most influential union of its kind globally.
Members of the actors’ union are voting on a newly negotiated agreement that was approved by the national board earlier this month, ahead of the current contract’s expiration at the end of June.
“The tone of the negotiation was much more collaborative and a lot more creativity was brought by both sides, so I really believe that the 2023 strikes — while they were very difficult for all of us — did help effectuate a reset in the relationship between the studios and the unions in general,” Crabtree-Ireland told AFP.
Approval would mean avoiding a repeat of disastrous 2023 strikes that shuttered productions, costing studios billions of dollars, while actors stood their ground against AI and other issues.
– AI’s evolutions –
The strike by the Screen Actors Guild-American Federation of Television and Radio Artists (SAG-AFTRA) lasted 118 days, with star-studded picket lines outside major studios in Los Angeles and New York, marking the longest such revolt in Hollywood history.
AI technology was a sticking point for many, and that tension persists, Crabtree-Ireland said.
“They do feel more secure than they did in 2023 but there’s still a very, very strong concern about AI — and especially because the generative AI tools have advanced so much in the last three years,” he said.
The latest agreement does not close the door on AI, but it does introduce new protections.
Under the new contract, digital replicas — which use AI or any technology to replicate an actual living or deceased performer — must “have informed consent and fair compensation,” Crabtree-Ireland said.
The contract allows for limited use of synthetics, under “unusual circumstances,” when a generative AI system can be used to create a character who is not based on any actual person in the world.
“There’s now process in place which would require the companies to come to the union if they want to use a synthetic in a project, they have to demonstrate to us that this synthetic brings a significant additional value to the production,” Crabtree-Ireland said.
“While this doesn’t rise quite to the level of a complete prohibition, it’s a very strong disincentive.”
Voting on the latest contract closes June 4.
UNM researchers use new machine learning method to detect self-harm history hidden in veterans’ medical records
University of New Mexico Health Sciences Center
Important mental health history is often present in medical records but hard to find, especially when it is missing from the diagnosis codes that clinicians, researchers and health systems use to search and count conditions.
A new study led by researchers at The University of New Mexico School of Medicine analyzed electronic health records for more than 1.3 million patients served by the Veterans Health Administration (VHA). Highlighting a common gap in how health systems track self-harm, the researchers found that diagnosis codes captured only about one-fourth of clinically documented self-harm history.
“For research and planning, if we only count what is easy to see in diagnosis codes, we may substantially underestimate the need for mental health services,” said Christophe Lambert, PhD, professor and interim chief of the Division of Translational Informatics in the UNM School of Medicine’s Department of Internal Medicine, and the study’s corresponding author. “Better measurement can help health systems plan better, help researchers study care more accurately and eventually help clinicians know when a patient may need a closer look.”
The study, published in the Journal of Medical Internet Research, used a novel machine learning method previously developed by members of the research team. Following expert chart review and statistical calibration, the researchers estimated that documented self-harm was present in about 7.9% of those patients seen by VHA clinicians – more than four times the 1.85% visible through diagnosis codes alone. The gap matters because missed history can affect clinical awareness, research findings and planning for mental health services.
Problem lists – the notations providers compile of their patients’ health conditions – showed another visibility gap. They are meant to flag important conditions for clinical teams, but in real-world care they are not always complete or consistently maintained. Among veterans with a diagnosis code for self-harm, 22.6% had self-harm or a history of self-harm listed on their VHA problem list. That means even when self-harm appeared in diagnosis codes, it was often missing from one of the record’s most visible summary fields.
Past self-harm is clinically important because it is one of the most important predictors of future self-harm and suicide risk. It can also shape how care is delivered, including how clinicians think about depression, PTSD, bipolar disorder, substance use, traumatic brain injury and other conditions that might occur alongside self-harm.
The authors note that VHA already uses specialized suicide and overdose reporting tools and does not rely only on diagnosis codes or problem lists to monitor suicide risk. This study looked at a different but related question: How much past self-harm history is visible in the parts of the record that researchers, care teams and health systems can most easily quantify and review at scale?
“This is a systems-level visibility problem,” Lambert said. “The record can be enormous. In our chart review, some patient records had more than 500,000 lines of notes. No clinician can be expected to read all of that during a normal visit.”
The study did not try to predict future self-harm or determine with certainty whether any one patient had self-harmed. Instead, the team tested whether a computer model could use patterns in structured electronic health record data to estimate the probability that self-harm history was present but missing from diagnosis codes, then compare those probabilities with expert review of clinical notes.
To do that, the team used a method called PULSNAR — Positive Unlabeled Learning Selected Not At Random, which was built for messy real-world health data. Most machine learning methods need clear examples of both “yes” and “no” cases. But in medical records, a missing diagnosis code does not prove that a patient never had the condition.
PULSNAR works with that uncertainty. It learns from patients who do have a code, then estimates how many similar patients might be present among those without a code. Its key advantage is that it does not assume coded cases are random and allows for the fact that some cases are more likely to be coded than others.
“Medical records can make self-harm hard to see in more than one way,” said Praveen Kumar, PhD, the study’s first author. “Sometimes the history is in a clinician’s note but not in the diagnosis codes. Other times, the record may contain risk factors, injuries, poisonings, or behaviors that are consistent with self-harm, even though the record alone does not prove what happened or why.
“Our method can help flag both patterns for review. This study could verify the first pattern, because the evidence was already in the notes. The second pattern may be just as important, but confirming it would require talking with patients or using information beyond the medical record.”
The research team included experts from the UNM Health Sciences Center, the Raymond G. Murphy Veterans Affairs (VA) Medical Center, Vanderbilt University Medical Center, the VA Tennessee Valley Healthcare System, the VA Office of Mental Health, Greer Black Company, and the UNM Department of Economics. The team brought together expertise in medical informatics, computer science, psychiatry, biomedical informatics, economics, statistics and health services research.
The self-harm study is part of a broader research program using positive-and-unlabeled learning to find conditions that may be under-recorded in standard medical data, the investigators said. The team has already published a related study using this approach to detect under-coded opioid use disorder, and ongoing work is extending it to other conditions where the medical record may not show the full picture, including unrecognized PTSD, depression, bipolar disorder and sleep disorders.
The method could complement broader VHA mental health and suicide-prevention efforts by adding a scalable way to measure conditions that may be under-recorded or hard to see in standard medical data. The investigators emphasized that the method is still a research tool and is not ready to be used by itself in clinical care, although with further development, it could help health systems better estimate under-recorded mental health conditions, find documented history that is not clearly visible, and identify records that may warrant closer review.
“Self-harm history matters too much to stay buried in records that are not practical to review line by line during routine care,” Lambert said. “Our work is about helping researchers and health systems find documented history and clinically relevant patterns in the data, so care teams can have a more complete picture of the people they serve.”
Journal
Journal of Medical Internet Research
DOI
10.2196/89071
Method of Research
Data/statistical analysis
Subject of Research
People
Article Title
Detecting Uncoded Self-Harm in Veterans’ Electronic Health Records Using Positive and Unlabeled Learning: Retrospective Cohort Study
Article Publication Date
4-Jun-2026
How Bristol researchers are using visual AI to improve wildlife conservation
University of Bristol
image:
Image of SA-FARI tracking squirrel monkeys
view moreCredit: SA-FARI
Wildlife research projects worldwide could benefit from a new AI system which can automatically find, name, and follow individual animals in footage.
A University of Bristol team working on Animal Biometrics and AI for Conservation have been key contributors to the SA-FARI (Segment Anything in Footage of Animals for Recognition and Identification) project, developed by an international consortium with first author Dante Wasmuht and senior author Didac Suris, overall led by ConservationX Labs (CXL) and META.
SA-FARI builds on META’s latest Segment Anything Model 3 (SAM3) which is a foundational and cutting-edge Vision-Language Model that is designed to use text and visual prompts to precisely identify, segment, and follow objects in images or videos.
This enables researchers to track animals in footage using ‘masklets’ which represent the exact outline of an animal in a video from frame-to-frame through time. It means the animal can be accurately separated from its background and form the basis of individual and behavioural analysis. This method has the potential to save thousands of hours for researchers using camera trap surveys in terms of viewing content manually.
The SA-FARI paper will be presented on Saturday 6 June at the Conference for Computer Vision and Pattern Recognition (CVPR) in Denver, USA, widely regarded as the leading conference for visual AI.
The paper has been selected as an Award Candidate at CVPR. For the Bristol team working on Animal Biometrics and AI Conservation, this marks a second consecutive year to get such a prestigious international nomination.
Tilo Burghardt, Professor of Computer Vision and Animal Biometrics at the University of Bristol, said: “Global problems require global solutions. Based on the group’s pioneering track record of over 20 years, the University of Bristol is regarded as one of the go-to places for using AI for conservation in the UK and beyond, and is an important part of a growing international community working in this area.”
Dr Otto Brookes, Lecturer in AI and Animal Biometrics from the Bristol team added: "The ability to locate animals in space and time is incredibly important for wildlife monitoring – it is a prerequisite for many tasks such as recognising behaviour and distinguishing individuals from one another and ultimately measuring how animals respond to conservation interventions."
The project trained and benchmarked an AI system which can automatically detect, name and track animals of nearly 100 species pixel-accurately in footage. To do this, a vast dataset of more than 11,000 wildlife videos taken in natural habitats was curated and annotated. The project offers this data freely downloadable for biologists, researchers, and conservationists to boost ecological projects worldwide with cutting-edge AI powers.
Professor Burghardt believes the SA-FARI work has the potential to be extended in the future by others by adding new features such as tracking animal body pose, depth and natural language descriptions.
The SA-FARI project was led by CXL and META, with important inputs from co-authors including Dr Otto Brookes and Prof Majid Mirmehdi from the University of Bristol, teams from the Hasso Plattner Institute, the University of Oviedo, Osa Conservation, the Senckenberg Museum of Natural History, the Max Planck Institute for Evolutionary Anthropology, and Climate Corridors – together the group pulled off a project of significant scale and truly inter-disciplinary reach.
Paper: “The SA-FARI Dataset: Segment Anything in Footage of Animals for Recognition and Identification”, by D. Francisco Wasmuht et al. CVPR 2026
Still of SA-FARI being used to track a Great Tinamou bird in footage
Credit
SA-FARI
Article Title
The SA-FARI Dataset: Segment Anything in Footage of Animals for Recognition and Identification
Article Publication Date
6-Jun-2026
New AI tools could help eye doctors diagnose retinal disease faster
Study finds that programs that analyze eye images could also speed up drug trials
Non-invasive eye scans allow doctors a zoomed-in, three-dimensional look beneath the eye’s surface without causing discomfort or pain to the patient. Used routinely in clinics worldwide, the scans produce detailed views of individual layers of the eye’s interior to help diagnose conditions that threaten vision. But with that level of precision comes a flood of data — hundreds of images per scan that physicians have to review manually, a time-consuming process that is vulnerable to human error.
Now, researchers at Washington University School of Medicine in St. Louis, in collaboration with colleagues at the University of Washington in Seattle and Genentech, Inc., have developed an experimental artificial intelligence (AI) system that can speed the scan review process and help doctors spot subtle signs of eye disease sooner. The technology, called OCTCube-M, includes a family of three AI models that are designed to read and interpret 3D images of the eye’s retina as well as other types of eye scans.
In a new study, the researchers found that, compared with older models, the new AI system more accurately identified eight different retinal diseases, including age-related macular degeneration, a common disease that damages the retina and is the leading cause of blindness in people over 50. It also was more accurate in its predictions of how fast a severe form of this condition, called geographic atrophy, would progress.
The findings describing the technology in its research stage were published recently in Nature Biomedical Engineering.
“Today’s eye scans provide physicians an unprecedented, highly detailed view of the inside of the eye, revealing structures and subtle changes that would otherwise go undetected,” said the study’s co-corresponding author Aaron Lee, MD, the Arthur W. Stickle Distinguished Professor of Ophthalmology and Visual Sciences and head of the John F. Hardesty, MD Department of Ophthalmology & Visual Sciences at WashU Medicine. “But we still lack the tools to help physicians process the volume of generated images. Our AI system has the potential to empower physicians to make faster diagnoses, tailor treatment more precisely and design clinical trials that bring new therapies to patients faster.”
Additionally, the study showed that the model could infer health risks beyond the eye, predicting outcomes such as heart attack, stroke and kidney failure based solely on retinal imaging. The tiny blood vessels in the retina are anatomically and developmentally the same as those in the kidney, and the processes that lead to plaque buildup inside the walls of blood vessels that feed the heart and brain also leave signatures in the eye.
“The model has the potential to turn a simple eye exam into a powerful tool for helping to detect illness beyond the eye,” said Lee. “It opens the door to earlier detection, more precise monitoring and potentially better outcomes for patients who might otherwise go undiagnosed until their disease is far more advanced.”
A diagnosis needle in a haystack of data
At least 2.2 billion people worldwide have vision impairment, according to the World Health Organization. The imaging method known as optical coherence tomography has transformed the diagnosis and care of conditions that cause vision loss by generating, via a single and swift scan, hundreds of cross-sectional images that together form a detailed, 3D picture of the retina and the optic nerve. It can reveal early signs of different eye diseases such as glaucoma, macular degeneration and diabetic retinopathy, among other conditions.
AI, meanwhile, has become a powerful tool for processing large medical datasets, and Lee and colleagues have made notable contributions to the field. Several years ago, they published, in Nature, results describing a model that is better at diagnosing eye disease in two-dimensional retinal images compared to older models.
Because the model was trained on 2D tomography images, the researchers sought to determine if adding 3D tomography images could further improve disease diagnosis and prognosis. Because disease often extends in all three dimensions around the fovea, a small pit in the center of the retina responsible for the sharp, detailed vision required to read text and recognize faces, they hypothesized that training models on 3D images would provide more complete and accurate views of the tissue. To that end, more than 26,000 3D optical coherence tomography images comprising 1.62 million individual retinal slices — cross-sectional images of the retina — were used to train OCTCube-M.
When compared to the model trained on 2D images, OCTCube-M more accurately identified six of the eight retinal diseases by about four to six percentage points. That translates to the tool finding 43 to 60 additional cases out of every 1,000 individuals with eye disease. This was true across scans taken from individuals at multiple clinical sites, imaging modalities and diverse patient populations.
The eight diseases identified by the model include serious conditions that primarily affect the back of the eye, including the retina and optic nerve. Together they are the leading causes of vision loss and are linked to other conditions such as diabetes, hypertension and cardiovascular disease.
The researchers, including Cecilia S. Lee, MD, the Jane Hardesty Poole Distinguished Professor in ophthalmology and visual sciences at WashU Medicine; Sheng Wang, PhD, an assistant professor in the Paul G. Allen School of Computer Science & Engineering at the University of Washington; and Miao Zhang, PhD, a senior AI scientist at San Francisco-based biotech company Genentech, then adapted the 3D model by adding data from two other eye imaging techniques — infrared retinal imaging and fundus autofluorescence imaging.
By combining optical coherence tomography with one or both of the other imaging types, the AI models can construct a more complete view of the eye and a deeper understanding of what’s happening inside, Aaron Lee explained. Indeed, the model trained on all three imaging types excelled at predicting the growth rate of the severe form of macular degeneration, geographic atrophy, significantly outperforming the current state-of-the art model that trains only on fundus autofluorescence images of the retina by an average of nearly 50%.
Geographic atrophy affects about 5 million people worldwide, and there are few effective treatment options. By providing information on the growth rate of the condition, Lee and colleagues’ tool could effectively detect and classify the stage of the illness, information that researchers could use to design better clinical trials of potential therapies for the disease.
“By better predicting how fast disease will worsen, we can run smaller, more efficient studies,” Lee said. “That could lower costs, shorten the time it takes to test new therapies, reduce the number of people exposed to treatments that don’t work and help effective drugs reach patients sooner.”
Next, the WashU Medicine researchers will train OCTCube-M with larger datasets encompassing more patients, more diseases and even more types of imaging data to continue improving upon it.
Journal
Nature Biomedical Engineering
Method of Research
Computational simulation/modeling
Subject of Research
People
Article Title
A 3D multi-modal foundation model for optical coherence tomography.
COI Statement
Zhang M and Steffen V are employees of Genentech, Inc. and hold stock in F. Hoffmann–La Roche Ltd. Cukras CA is an employee of F. Hoffmann–La Roche Ltd. and holds stock in the company.
