Decoding Neuronal Activity ‘Streams’ in Cultured Networks: A Transfer Learning Approach Using VGG16 and Signal-to-Image Transformation

Decoding stimulus-evoked electrical activity in cultured neuronal networks is essential for elucidating the fundamental mechanisms of brain function. Using a microelectrode array, we captured the spatiotemporal dynamics of neuronal population activity and developed an analytical framework based on stream representation of neuronal activity propagation. From the reconstructed instantaneous spatial patterns (ISPs), we generated two types of images: time-information-prioritized neuronal activity pattern (TIP-NAP) images and space-information-prioritized neuronal activity pattern (SIP-NAP) images. Each image type was trained and classified using transfer learning with the VGG-16 convolutional neural network. The results showed that the classification accuracy of TIP-NAP images (99.9 ± 0.5 %) significantly outperformed that of SIP-NAP images (96.3 ± 9.1%) and the conventional SVM-based method (97.3–97.7%). The primary contributions of this study are threefold: First, we established that our proposed Signal-to-Image Transformation technique, based on ISPs, effectively represents the stream of neuronal activity. Second, we demonstrated that converting TIP-NAP images enables highly accurate decoding via VGG-16 transfer learning. Third, our comparative analysis provides compelling evidence that temporal continuity—the temporal structure of neuronal activity—plays a more critical role in information encoding than instantaneous spatial synchronization within biological networks.