Kaneiya Desai¹ (), Bilge Sayim², Dennis M. Levi¹, Ângela Gomes Tomaz¹; ¹University of California, Berkeley, ²University of Lille, France

Amblyopia is a neurodevelopmental disorder characterized by reduced visual acuity due to uncorrelated visual input early in life. Observers with amblyopia perceive stimuli as distorted (e.g., straight lines perceived as jagged) and experience stronger foveal crowding (i.e., worse performance with flanked targets) than controls. In a previous study (Gomes Tomaz et al., 2023), observers recreated the appearance of letters presented isolated or flanked to the fovea of the dominant/fellow or non-dominant/amblyopic eye. Here, we trained a convolutional neural network (CNN) with the appearance captured in that study to predict whether a response was made by an observer with or without amblyopia and under which condition, isolated or flanked. The architecture consisted of four convolutional, activation, and maximum pooling layers each, followed by two fully connected layers. Batch normalization followed the first, third, and fourth activation layers. The training dataset consisted of all appearance responses (captures), along with their classification label (e.g., “control” for a response made by a control observer (when classifying for group)). The validation dataset, consisting of 30% of all appearance captures, was randomized and balanced within and between all variables. The captures included in the validation dataset were not part of the training dataset. The best-performing models for classifying group and condition showed training accuracies of 88.0% and 95.0% and validation accuracies of 69.2% and 90.4%, respectively. The lower accuracy of the group predicting model is likely due to the high variability of features in the appearance space that characterizes each group. These results highlight the potential of CNNs in the analysis and classification of target appearance. Accurate models that classify and predict stimulus appearance in amblyopia will be lesioned to further study amblyopic visual perception in the future.

Acknowledgements: Supported by a grant from the National Eye Institute awarded to Dennis Levi (R21EY030609)