Recognition of ex vivo spike response of retina using a convolution-attention combined framework

This study aimed to develop and evaluate a compact spike decoding framework for classifying visual stimulus from retinal MEA recordings that may have the problem of poor electrode-tissue coupling. Specifically, we sought to establish a convolution-attention framework that can extract spike features and capture longer-range temporal dependencies in visually evoked retinal activities. Approach: MEA recordings from ex vivo mouse retinas were acquired under controlled colour and shape stimulation paradigms. Spike trains were converted into spike-count sequences and used to train an artificial neural network framework combining convolutional neural network layers with temporal multi-head attention modules. Convolutional layers extracted local temporal features, while the attention modules facilitated broader temporal relationships across the response sequence. Main results: The proposed convolution-attention framework achieved good performance across both colour and shape tasks. The proposed framework reached 93.250% accuracy in colour classification and 73.161% accuracy in the shape classification task. These results show that the proposed framework can decode visually evoked retinal spike responses under different stimulus conditions and retains stable performance across tasks with different levels of complexity. Significance: This study presents a compact decoding framework for retinal spike trains and provides a potential computational backbone for spike-based neural decoding in bio-electronic interfaces. By improving stimulus classification from retinal neural activity, the proposed model may provide a cornerstone to support future visual neuroengineering and bio-digital convergence. Keywords: ex vivo retina, MEA, spike classification, deep learning framework.