Ning-Ting (Michelyne) Hsu¹, Tiffany Tai¹, Haley G Frey², Katrina A Wolters², Shao-Min (Sean) Hung³, Po-Jang (Brown) Hsieh¹, David Whitney²; ¹Department of Psychology, National Taiwan University, ²University of California, Berkeley, ³Waseda Institute for Advanced Study, Waseda University

Dermatologists demonstrate remarkable efficiency when examining large areas of patients' skin, such as the arms or back, spending only brief glances at individual spots. This efficiency suggests they may use ensemble perception, a cognitive process in which clinicians extract overall statistical patterns from groups of lesions rather than examining each one individually. Understanding whether this mechanism operates in dermatology is crucial, as it would reveal how clinicians achieve rapid, wide-field anomaly detection under limited time. Currently, no direct evidence confirms that ensemble perception applies to skin lesion detection, representing a significant gap in our understanding. If dermatologists rely on ensemble processing, it will indicate that successful diagnosis depends not just on individual lesion features but on how a lesion compares to the statistical structure of the whole visual field. However, this introduces specific biases—malignant lesions may appear more benign when surrounded by benign ones, and higher variability may reduce sensitivity to malignancy. Contrarily, if ensemble perception is weak, clinicians would rely on serial search, potentially reshaping how diagnostic errors are understood. The research test whether observers can perceive ensemble characteristics of skin lesion malignancy, bridging vision science with real-world cancer detection. We recruited participants without dermatological training. Seven experimental conditions included both benign and malignant lesion images from the ISIC Challenge Datasets 2019, with malignant images ranging from zero to six. Each trial randomly assigned set sizes of one, two, four, or six images. After exposure, participants judged average malignancy and confidence using a continuous rating slider. Results showed a strong positive correlation (r = 0.64, p < 0.001) between participant responses (N = 6) and average actual malignancy ratings, demonstrating that observers can use ensemble coding to perceive malignancy during rapid screening, opening important questions about training this mechanism and diminishing its inherent biases.

Acknowledgements: supported in part by NIH R01CA236793