Study finds pigeons can learn to spot abnormalities on lung CTs
Bottom line
Pigeons can be trained to distinguish normal from abnormal lung CT scans, according to a new paper in Animal Cognition that tested whether the birds could spot solid lung nodules in short CT “movies.” In the study, six of eight pigeons learned the task, then generalized that learning to previously unseen scans, suggesting they weren’t just memorizing images. The discrimination also transferred to visually different abnormalities, including ground-glass nodules and emphysema, which the authors say points to broader pattern-recognition abilities rather than a single simple cue. The researchers frame the work not as a replacement for radiologists, but as a model for studying how biological vision systems detect subtle abnormalities and how that insight might inform imaging AI and training. (link.springer.com)
Why it matters: For veterinary professionals, the study is less about pigeons diagnosing patients and more about how visual expertise develops in image interpretation. The authors note that lung nodule detection is a difficult task even for trained human readers, and they argue that understanding the implicit visual features pigeons use could help researchers refine diagnostic imaging workflows, observer training, or AI tools. That’s relevant in veterinary radiology, too, where clinicians increasingly work alongside pattern-recognition software and still need to understand where human and machine perception can miss subtle disease. Earlier work from the same research area also found pigeons could be trained on breast pathology and radiology images, reinforcing the idea that animal visual systems may offer an unusual but useful test bed for imaging research. (link.springer.com)
What to watch: The next step is whether this line of research can identify the specific visual cues pigeons are using, and whether those cues can be translated into better radiology training or more reliable diagnostic AI. (link.springer.com)
Key facts
- Study
- Pigeons were trained to classify short chest CT sequences as normal or abnormal.
- Journal
- Animal Cognition
- Training target
- Solid lung nodules.
- Sample size
- Eight pigeons.
- Result
- Six pigeons learned the discrimination.
- Transfer finding
- Performance generalized to previously unseen scans.
- Additional abnormalities
- Ground-glass nodules and emphysema.
- CT format
- 30-frame sequences over 10 seconds.
A new Animal Cognition study suggests pigeons are better at reading some medical images than many people might expect. Researchers trained the birds to classify short chest CT sequences as normal or abnormal based on the presence of solid lung nodules, then tested whether the birds could carry that learning to new scans. Six pigeons learned the discrimination, and their performance generalized to novel images, a key sign that they were recognizing patterns rather than memorizing a fixed set of pictures. (link.springer.com)
The work builds on a longer-running line of research into pigeons as models of visual categorization. A 2015 study found pigeons could learn to distinguish benign from malignant breast histopathology images and some radiology images, though performance varied by image type. A later commentary argued that the real value of those experiments was not clinical deployment, but what they reveal about perception, observer training, and the evaluation of imaging systems. The new CT study extends that concept into dynamic imaging, where the observer must integrate information across multiple slices over time. (pmc.ncbi.nlm.nih.gov)
In the new experiment, the birds viewed 30-frame CT sequences over 10 seconds. Five pigeons were rewarded for responding to abnormal scans, while three were rewarded for responding to normal ones. The CT cases came from a custom lung dataset that had been confirmed by clinical report, over-read by two thoracic radiologists, and checked with an AI-based nodule detection algorithm. According to the paper, the nodules appeared in four to seven slices on average, forcing the birds to process changes across space and time rather than rely on a single static image. (link.springer.com)
The most interesting result may be the transfer test. After learning to classify solid lung nodules, the pigeons also responded appropriately to previously unseen abnormalities, including ground-glass nodules and emphysema. The authors say that matters because those abnormalities do not share the same obvious appearance-and-disappearance pattern as the original training set. In other words, the birds may have picked up on a broader perceptual signature of abnormal lung tissue, even if researchers still don’t fully understand what that signature is. (link.springer.com)
Industry and expert reaction around this research area has tended to be cautious but intrigued. Prior commentary in the pathology literature has argued that pigeons are best understood as surrogate observers for studying image perception, not as practical diagnostic replacements. That distinction is important: the point is not to put birds in the reading room, but to use a nonhuman visual system to test how image features, display choices, or compression methods affect detectability. The new paper makes a similar argument, positioning pigeons as a tool to study implicit visual cognition in ways that could eventually support human readers and imaging developers. (sciencedirect.com)
Why it matters: For veterinary professionals, especially those in diagnostic imaging, oncology, and academic medicine, the study is a reminder that image interpretation depends on more than explicit training rules. The authors note that lung nodule detection remains error-prone for humans, and that AI systems also miss abnormalities or generate false positives. In veterinary settings, where imaging volumes are growing and AI adoption is still uneven, research like this adds to a broader conversation about how clinicians learn visual pattern recognition, how algorithms should be validated, and how subtle abnormalities might be surfaced more reliably. It also underscores a practical point for veterinary teams: better diagnostics may come not only from new software, but from a deeper understanding of perception itself. (link.springer.com)
There’s also a comparative medicine angle. While the study used human CT data, its implications cross species because the underlying challenge, detecting faint, variable abnormalities in complex images, is central to veterinary radiology as well. For clinicians who interpret thoracic imaging in dogs, cats, horses, or exotics, the paper offers a useful conceptual parallel: even very different biological visual systems can extract meaningful diagnostic patterns, which may help researchers think differently about training, second-reader systems, and human-AI collaboration. This is especially relevant as veterinary medicine continues to adopt advanced imaging and computer-assisted review tools. (link.springer.com)
What to watch: The next phase will likely focus on defining exactly which visual features the pigeons used, testing whether those features align with or differ from human expert strategies, and determining whether the findings can improve radiologist education or AI model design. If that translation happens, the lasting value of this work may be methodological, not novelty-driven: a new way to study how diagnostic eyes, human or otherwise, learn to see disease. (link.springer.com)