Diagnostic features predicted by deep learning improve human object recognition in simulated prosthetic vision.

Objective. Cortical visual prostheses aim to partially restore vision by electrically stimulating V1. Yet, their effectiveness is constrained by the limited number of simultaneously elicitable phosphenes. Identifying sparse but informative visual features for object recognition is therefore crucial to enable functional prosthetic vision. In this study, we aim to target diagnostic features for object recognition.Approach. We combined a deep learning pipeline with bioplausible phosphene simulation to optimize phosphene placement for object recognition. The phosphene renderings were presented to 63 participants who performed an object recognition task under simulated prosthetic vision with varying electrode densities.Main results. Performance in the proposed deep neural network (DNN) phosphene placement strategy was compared to contour-based and luminance-based phosphene placements. Results demonstrate that the DNN-based approach significantly reduces the number of phosphenes required for object recognition. Importantly, this advantage was especially pronounced at low level electrode densities.Significance. These findings indicate that DNNs can effectively identify diagnostic object features and enhance recognition under simulated prosthetic vision. The proposed pipeline provides a principled method for optimizing visual representations in cortical visual prostheses, highlighting the importance of task-relevant feature selection imposed by constraints on simultaneous stimulation capacity.