But a study in Scientific Reports highlights a hidden challenge of using AI in medical imaging research — the phenomenon of highly accurate yet potentially misleading results known as “shortcut learning.”

The researchers analyzed more than 25,000 knee X-rays from the National Institutes of Health-funded Osteoarthritis Initiative and found that AI models can “predict” unrelated and implausible traits such as whether patients abstained from eating refried beans or beer. While these predictions have no medical basis, the models achieved surprising levels of accuracy by exploiting subtle and unintended patterns in the data.

“While AI has the potential to transform medical imaging, we must be cautious,” says the study’s senior author, Dr. Peter Schilling, an orthopaedic surgeon at Dartmouth Health’s Dartmouth Hitchcock Medical Center and an assistant professor of orthopaedics in Dartmouth’s Geisel School of Medicine.

“These models can see patterns humans cannot, but not all patterns they identify are meaningful or reliable,” Schilling says. “It’s crucial to recognize these risks to prevent misleading conclusions and ensure scientific integrity.”

The researchers examined how AI algorithms often rely on confounding variables — such as differences in X-ray equipment or clinical site markers — to make predictions rather than medically meaningful features. Attempts to eliminate these biases were only marginally successful — the AI models would just “learn” other hidden data patterns.

“This goes beyond bias from clues of race or gender,” says Brandon Hill, a co-author of the study and a machine learning scientist at Dartmouth Hitchcock. “We found the algorithm could even learn to predict the year an X-ray was taken. It’s pernicious — when you prevent it from learning one of these elements, it will instead learn another it previously ignored. This danger can lead to some really dodgy claims, and researchers need to be aware of how readily this happens when using this technique.”

The findings underscore the need for rigorous evaluation standards in AI-based medical research. Overreliance on standard algorithms without deeper scrutiny could lead to erroneous clinical insights and treatment pathways.

“The burden of proof just goes way up when it comes to using models for the discovery of new patterns in medicine,” Hill says. “Part of the problem is our own bias. It is incredibly easy to fall into the trap of presuming that the model ‘sees’ the same way we do. In the end, it doesn’t.”

“AI is almost like dealing with an alien intelligence,” Hill continues. “You want to say the model is ‘cheating,’ but that anthropomorphizes the technology. It learned a way to solve the task given to it, but not necessarily how a person would. It doesn’t have logic or reasoning as we typically understand it.”

Schilling, Hill, and study co-author Frances Koback, a third-year medical student in Dartmouth’s Geisel School, conducted the study in collaboration with the Veterans Affairs Medical Center in White River Junction, Vt.

Where does all that energy come from? For many years, scientists believed it came from shocks that occur in extreme astrophysical environments — when, for example, a star explodes before forming a black hole, causing a huge explosion that kicks up particles.

That theory was plausible, but, according to new research published this week in The Astrophysical Journal Letters, the observations are better explained by a different mechanism. The source of the cosmic rays’ energy, the researchers found, is more likely magnetic turbulence. The paper’s authors found that magnetic fields in these environments tangle and turn, rapidly accelerating particles and sharply increasing their energy up to an abrupt cutoff.

“These findings help solve enduring questions that are of great interest to both astrophysicists and particle physicists about how these cosmic rays get their energy,” said Luca Comisso, associate research scientist in the Columbia Astrophysics Lab, and one of the paper’s authors.

The paper complements research published last year by Comisso and collaborators on the sun’s energetic particles, which they also found emerge from magnetic fields in the sun’s corona. In that paper, Comisso and his colleagues discovered ways to better predict where those energetic particles would emerge.

Ultra-high energy cosmic rays are orders of magnitude more powerful than the sun’s energetic particles: They can reach up to 10²⁰ electron volts, whereas particles from the Sun can reach up to 10¹⁰ electron volts, a 10-order-of-magnitude difference. (To give an idea of this vast difference in scale, consider the difference in weight between a grain of rice with a mass of about 0.05 grams and a 500-ton Airbus A380, the world’s largest passenger aircraft.) “It’s interesting that these two extremely different environments share something in common: their magnetic fields are highly tangled and this tangled nature is crucial for energizing particles,” Comisso said.

“Remarkably, the data on ultra-high energy cosmic rays clearly prefers the predictions of magnetic turbulence over those of shock acceleration. This is a real breakthrough for the field,” said Glennys R. Farrar, an author on the paper and professor of physics at New York University.

The research was supported by the National Science Foundation.

Kirtland’s warblers have also been known as jack pine warblers, as their breeding habitat is limited to young jack-pine forests in Michigan. The species nearly went extinct in the 1970s due to threatened habitat and the population decline was exacerbated by brown-headed cowbirds, who lay their eggs in the nests of other bird species — including warblers — and divert resources away from their chicks. Kirtland’s warblers were classified as endangered under the Endangered Species Act of 1973. Thanks to intensive, targeted management over several decades, Kirtland’s warbler populations have recovered. The species was “delisted” — no longer considered endangered — in 2019, though management efforts remain critical to their survival, according to the researchers.

“When any population goes through a period when few breeding individuals are present, called a ‘bottleneck,’ there is always a chance of inbreeding, which can lead to reduced survival of future generations,” said Anna María Calderón, graduate student in biology in the Penn State Eberly College of Science and first author of the paper. “The 2019 decision to delist Kirtland’s warblers referenced a study that did not find strong evidence of inbreeding, but the tests available at the time were not very high resolution. We used updated sequencing technology to get a clearer picture of the genetic diversity and potential for inbreeding among these songbirds.”

Like other animals, birds receive one copy of their genome from each parent. Some genetic variants can have negative impacts on an animal’s survival or reproductive success. The likelihood of receiving the same potentially damaging variant from both parents increases with inbreeding, when genetically similar individuals mate.

Previous tests of genetic diversity — such as those referenced in the delisting decision — often focus on a small set of specific markers on the genome called microsatellites, which only provide part of the picture. However, advancements in sequencing technology now allow inexpensive analysis of the entire genome from multiple individuals.

“If you think of the genome as a movie of the blueprint of life, then using microsatellites is like trying to get the plot from half a dozen snapshots, while sequencing the entire genome is like watching an entire 4K-resolution film,” said David Toews, Louis Martarano Career Development Professor of Biology at Penn State and co-leader of the research team. “There’s so much more information to be obtained from sequencing the entire genome, and modern technology has made that analysis possible.”

The researchers sequenced the whole genome of Kirtland’s warblers and, as a point of comparison, they also sequenced whole genomes of two closely related species whose populations have remained large and stable, Hooded warblers and American redstarts. The team specifically looked at a measure of inbreeding called “runs of homozygosity.”

A given spot on the genome is called homozygous when the genetic information obtained from an animal’s maternal parent is the same as the information obtained from its paternal parent. Runs of homozygosity (ROH) are long, connected stretches of homozygosity, sometimes millions of letters of the genetic alphabet in length. Particularly long ROH indicate that the individuals who carry them are products of recent inbreeding, when genetically similar individuals mated and were thus more likely to pass on the same genetic information. Fewer and shorter stretches indicate that an individual’s parents are more distantly related.

“Runs of homozygosity give us a unique look into the past and can be an indicator of the genomic health of the population,” said Zachary Szpiech, assistant professor of biology at Penn State and the other co-leader of the research team. “For example, long ROH have been associated with how well other species survive in their first year. ROH can also flag potentially damaging genetic variants, which may be useful when considering the conservation of the bird.”

The researchers found exceptionally long ROH in Kirtland’s warblers, indicating very recent inbreeding that had not previously been identified in this species. Additionally, they found many small to medium-sized ROH in Kirtland’s warblers, while the two closely related species had almost none.

“The contrast between Kirtland’s warblers and their closest related species, which haven’t undergone any population bottlenecks, couldn’t be more stark,” Toews said. “We found no evidence of inbreeding whatsoever in Hooded warblers or American redstarts, and in comparison, the level of inbreeding for the Kirtland’s warblers is almost off the charts. We also found a high frequency of potentially damaging genetic variants. This gives us a clear picture of how the demographic history of songbirds can shape their genetic diversity.”

The researchers noted that the Kirtland’s warblers they sampled did not appear to have any physical deformities. However, they said that inbreeding could manifest during the stages of life they did not observe, for example during development or the energy-intensive task of migration, or impact aspects of their reproductive success, such as how many eggs they lay or how many hatch. According to Toews, continued monitoring of these birds to clarify the impacts of inbreeding will be critical to future conservation efforts.

“An open question in this species has been whether the population of Kirtland’s warblers has always been small, or if it was once large and crashed,” Calderón said. “We can actually use ROH to look into the past, using rates of recombination — one way genomes can swap information around — to estimate when certain combinations of genetic information appeared. From this, we can see that most of the DNA segments within the short ROH in these birds originated between 1874 and 1954, mostly before known bottlenecks. This suggests that Kirtland’s warblers may have always had small population sizes. This makes sense given the high specificity of these birds to young jack-pine forests.”

Comparing these results with information about the distribution and abundance of the species, which has been collected since the 1940s, can provide insight into the events that shaped the genomes of these birds. For example, the researchers dated the origin of many DNA segments within long ROH between 1940 to 1981. This is consistent with a sudden population collapse beginning in the early 1940s, as this bottleneck would result in more, longer ROH. The researchers also plan to sequence the genomes of museum samples, including birds collected as early as the late 1800s, to provide additional context around genetically important events.

“Although Kirtland’s warblers are celebrated as a conservation success story, they are highly inbred and have a high frequency of potentially damaging genetic variants that could impact their ability to persist,” Calderón said. “Our work serves as an important example of how measures of genetic health may tell a different story than simple population numbers. Both need to be considered when assessing a population’s recovery and when making conservation decisions.”

In addition to Calderón, Toews and Szpiech, the research team also includes Andrew Wood, a research technologist in Toews’ lab at the time of the research and now a postdoctoral researcher at the University of Minnesota.

Funding from the U.S. National Science Foundation Division of Environmental Biology, the Penn State Huck Institutes of the Life Sciences and the Penn State Eberly College of Science supported this research.

Unlike the typical quantum chip design that lays the information-processing qubits onto a 2-D grid, the team from the Cleland Lab has designed a modular quantum processor comprising a reconfigurable router as a central hub. This enables any two qubits to connect and entangle, where in the older system, qubits can only talk to the qubits physically nearest to them.

“A quantum computer won’t necessarily compete with a classical computer in things like memory size or CPU size,” said UChicago PME Prof. Andrew Cleland. “Instead, they take advantage of a fundamentally different scaling: Doubling a classical computer’s computational power requires twice as big a CPU, or twice the clock speed. Doubling a quantum computer only requires one additional qubit.”

Taking inspiration from classical computers, the design clusters qubits around a central router, similar to how PCs talk to each other through a central network hub. Quantum “switches” can connect and disconnect any qubit within a few nanoseconds, enabling high-fidelity quantum gates and the generation of quantum entanglement, a fundamental resource for quantum computing and communication.

“In principle there’s no limit to the number of qubits that can connect via the routers,” said UChicago PME PhD candidate Xuntao Wu. “You can connect more qubits if you want more processing power, as long as they fit in a certain footprint.”

Wu is the first author of a new paper published in Physical Review X that describes this new way of connecting superconducting qubits. The researchers’ new quantum chip is flexible, scalable and as modular as the chips in cellphones and laptops.

“Imagine you have a classical computer that has a motherboard integrating lots of different components, like your CPU or GPU, memory and other elements,” said Wu. “Part of our goal is to transfer this concept to the quantum realm.”

Size and Noise

Quantum computers are highly advanced yet delicate devices with the potential to transform fields such as telecommunications, healthcare, clean energy, and cryptography. Two things must happen before quantum computers can tackle these global problems to their fullest potential.

First, they must be scaled to large enough size with flexible operability.

“This scaling can offer solutions to computational problems that a classical computer simply cannot hope to solve, like factoring huge numbers and thereby cracking encryption codes,” Cleland said.

Second, they must be fault-tolerant, able to perform massive calculations with few errors, ideally surpassing the processing power of current state-of-the-art classical computers. The superconducting qubit platform, under development here, is one promising approach to building a quantum computer.

“A typical superconducting processor chip is a square shape with all the quantum bits fabricated on that. It’s a solid-state system on a planar structure,” said co-author Haoxiong Yan, who graduated from UChicago PME in the spring and now works as a quantum engineer for Applied Materials. “If you can imagine a 2-D array, like a square lattice, that’s the topology of typical superconducting quantum processors.”

Limitations in Typical Design

This typical design causes several limitations.

First, putting qubits on a grid means each qubit can only interact with, at most, four other qubits — its immediate neighbors to the north, south, east and west. Greater qubit connectivity usually enables a more powerful processor with respect to both flexibility and component overhead, but the four-neighbor limit is generally considered inherent to the planar design. This means for practical quantum computing applications, scaling the device using brutal force will likely result in unrealistic resource requirements.

Second, the nearest-neighbor connections will in turn limit the classes of quantum dynamics that can be implemented as well as the extent of parallelism the processor is able to execute.

Finally, if all qubits are fabricated on the same planar substrate, then this poses a significant challenge to the fabrication yield, as even a small number of failed devices means the processor won’t work.

“To undertake practical quantum computing, we need millions or even billions of qubits and we need to make everything perfectly,” Yan said.

Rethinking the Chip

To work around these issues, the team retouched the design of the quantum processor. The processor is designed to be modular, in a way that different components can be pre-selected before being mounted onto the processor motherboard.

The team’s next steps are working on ways to scale up the quantum processor to more qubits, find novel protocols for expanding the processor’s capabilities, and, potentially, find ways to link router-connected qubit clusters the way supercomputers link their component processors.

They’re also looking to expand the distance over which they can entangle qubits.

“Right now, the coupling range is sort of medium-range, on the order of millimeters,” Wu said. “So if we’re trying to think of ways to connect remote qubits, then we must explore new ways to integrate other kind of technologies with our current setup.”

Funding: Devices and experiments were supported by the Army Research Office and Laboratory for Physical Sciences (ARO Grant No. W911NF2310077) and by the Air Force Office of Scientific Research (AFOSR Grant No. FA9550-20-1-0270)