Continuous Prediction of Mice Lever-Pressing Kinematic Parameters by Background Removed Single-photon Calcium Images

Calcium imaging has gained extensive application in neural decoding tasks because of its high precision in observing cortical neural activity. Nevertheless, the immense data volume and complexity of automated signal extraction algorithms in calcium imaging result in significant delays in extracting neuronal calcium fluorescence signals, greatly constraining the efficiency of neural decoding research and its applicability in real-time tasks. Although a few studies have successfully used partial neuronal signals from calcium imaging data for real-time neural decoding and brain-computer interface tasks, they fail to leverage the complete neuronal dataset from experiments, which limits their ability to decode continuous and complex movements. In response to this challenge, we introduce a neural decoding method based on background-removed single-photon calcium images. This approach extracts three-dimensional spatiotemporal representations of neuronal activity via background removal and employs a decoder combining 3D-ResNet and RNN networks to enable continuous and rapid decoding of mouse lever-pressing kinematic parameters. Compared with traditional methods for neural decoding using single-photon calcium imaging, this approach offers higher accuracy and faster speed. Combined with real-time motion correction algorithms, the proposed neural decoding approach meets real-time decoding requirements at a 20Hz acquisition frame rate, achieving single decoding in just 21.8ms. This advancement significantly improves the efficiency of single-photon calcium imaging-based neural decoding, offering solutions for its application in real-time tasks, such as optical brain-computer interfaces.