Abstract
"Deep Learning" is a form of perceptual learning in which the trainee learns to perform a given task by learning the informative, often abstract, statistical patterns in the data from a relatively large set of labeled examples. We have previously reported that, using deep learning, naïve, non-professional human observers can be trained to detect camouflaged objects in natural scenes, or anomalies in radiological images (Chen and Hegdé, Psychol Sci 2014; Hegdé, J Vis 2014). By systematically manipulating the deep visual patterns (e.g., principal components [PCs]) using image synthesis algorithms, we have identified the patterns that such non-professional 'experts' use in detecting cancers in screening mammograms. But it is not known whether or to what extent practicing radiologists can or do use the same patterns. To help address this issue, we tested practicing radiologists (N = 9; 3 mammography specialists) under comparable conditions. Briefly, either original mammograms or synthesized counterparts that were missing 0 to 2 of the previously characterized PCs were viewed ad libitum one per trial. Depending on the trial, subjects indicated whether the mammogram contained a cancer (detection task), or whether the image was original or synthesized (discrimination task). Subjects were unable to discriminate original vs. synthesized images when the latter contained all PCs (d' = 0.38, p > 0.05), indicating that the two sets of images were mutually perceptually metameric. In the detection task, the performance of the radiologists covaried with the cumulative eigenvalue of the PCs in the image and with that of the non-professional subjects (two-way ANCOVA, eigenvalue x training mode; p < 0.05 for both factors and interaction). Together, our results indicate that at least some of the visual patterns used by professionally trained radiologists are the same as that learned and used by non-professionals trained in the laboratory.
Meeting abstract presented at VSS 2017