Researchers Use Pigeons To Train Cancer-Detection AI

According to reporting by Popular Science and Economic Times, and the peer-reviewed paper in Animal Cognition (Springer, February 2026), Dr. Gregory DiGirolamo of the College of the Holy Cross led a study in which six pigeons were trained to watch short CT scan videos and indicate whether a lung nodule was present. Economic Times reports that some birds received food rewards for correctly identifying scans with nodules, while others were rewarded for correctly recognising normal scans, and the pigeons were able to generalize to previously unseen scans. Economic Times reports the birds also detected other lung abnormalities, including emphysema and ground-glass nodules.

Popular Science reports that in 2025 DiGirolamo and colleagues published an eye-tracking study showing that radiologists often fixate on suspicious lung nodules and exhibit pupil dilation even when they later mark a scan as normal. Popular Science frames the pigeon experiments as a way to study that nonconscious visual signal without human conscious decision processes interfering.

Studies that use nonhuman visual systems to probe perception provide alternative supervisory signals that differ from explicit human labels. For practitioners: animal vision can offer examples of invariances, salience, and pattern recognition not captured by standard labeled datasets, and those signals can be encoded as auxiliary training targets, attention priors, or contrastive tasks when developing medical imaging models.

Observers place this work at the intersection of perceptual science and model training. For practitioners: leveraging behavioral readouts such as fixation patterns, reward-conditioned responses, or other proxy labels can supplement scarce clinical labels and help surface subtle features that standard annotation workflows miss. This is especially relevant in medical imaging where early-stage abnormalities are rare and hard to label consistently.

whether signals derived from pigeon behavior can be translated into reproducible algorithmic training signals, whether such signals improve clinical metrics like sensitivity at fixed specificity, and whether methods are validated on larger clinical datasets. Observers will watch for peer-reviewed follow-on work and methodological details specifying how animal-derived signals are converted into model objectives. Popular Science reports DiGirolamo plans to use eye gaze-tracking and physiology data (such as pupil widening) to capture how radiologists respond to subtle abnormalities and feed those patterns into AI models.