Learning to perceive the gist of cancer through perceptual training

Rapid extraction of global structural regularities provides us with general information about our visual environment, the so-called gist. Similarly, medical experts can detect the gist of abnormality from medical images. We investigated whether naïve observers could learn the gist of medical abnormality. Fifteen participants completed nine training sessions (=6480 trials), receiving categorical feedback (cancer present/developing or absent), where viewing time started at 2500 ms and gradually reduced towards 500 ms based on performance. They viewed unilateral mammograms of four categories: normal, obvious-abnormal, subtle-abnormal, and global signals of abnormality (mammograms with no visible lesions but from breasts contralateral to or years prior to development of cancer). Performance was measured on a 200-mammogram test-set viewed for 500 ms without feedback pre-training, post-training, and at retention (7-10 days). D’ was modulated by category (p=.001, ηp2=.41), with a trending session effect (p=.068, ηp2=.18), without interaction (p=.301, ηp2=.08). Log linear likelihood-derived AUCs show that 11 participants performed above chance after training but only 7 at retention. Sub-analysis of 9 learners (positive correlation of d’ across training) showed a significant session (p=.026, ηp2=.37) and category (p<.001, ηp2=.62) effect, without interaction (p=.924, ηp2=.03). While naïve observers showed higher performance for mammograms with obvious and subtle lesions, critically they learned to detect abnormal gist in mammograms without any localizable lesion. Interestingly, while a state-of-the-art breast cancer classifier recognised obvious and subtle lesions as malignant, it struggled with the global category, suggesting observers learned a complex global signal not easily captured by current DNNs. Gist of abnormality can be learned through perceptual/incidental learning in mammograms both with and without visible abnormalities, although there were individual differences. Gist was poorly retained, suggesting that perceptual tuning to gist is lost when not used converging with findings that radiologists’ gist performance correlates with cases reviewed in a year, not years of experience.