Institution: University of California Berkeley
Poster Title: Harnessing convolutional neural networks (CNNs): Predicting amblyopia from the appearance of isolated and crowded stimuli
Abstract: Amblyopia is a neurodevelopmental disorder characterized by reduced visual acuity due to different visual input from the two eyes early in life. Previous research has shown that observers with amblyopia have stronger foveal crowding (i.e., worse performance with flanked targets) than neurotypical controls and perceive stimuli as distorted (e.g., straight line elements are perceived as jagged) and lower in contrast. In a previous study (Gomes Tomaz et al., 2023), observers were presented with isolated and flanked letter stimuli to the fovea of the dominant /fellow, or non-dominant/amblyopic eye. Observers recreated stimulus appearance by selecting squares on a 9x9 square-grid interface. Here, we trained a convolutional neural network (CNN) with the dataset of observersÂ’ depictions to find the best models to predict whether a depiction was made by an observer with or without amblyopia, and to an isolated or flanked stimulus. The CNN architecture consisted of 3 convolutional, activation, and maximum pooling layers, followed by fully connected layers to classify for group (control, amblyopia) and crowding (isolated, flanked) variables. The model architecture was trained 10 times. The validation dataset was randomized, but balanced between all variables present in the dataset to ensure a representative sample of the appearance space. The most accurate models for classifying between groups (control, amblyopia) and crowding condition (isolated, flanked) have training accuracies of 74.4% and 97.2% and validation accuracies of 68.2% and 93.8%, respectively. The current results indicate that CNNs are useful in classifying images of target depictions. Lower accuracy for the group prediction model is potentially due to high variability in the appearance space that characterizes each group. In the future, accurate models that classify and predict stimulus appearance in amblyopia will be lesioned to further study amblyopic visual perception.