Fig 1.
Block diagram of the real-time spike sorting system.
The system is comprised of a desktop computer and an FPGA module. The system can measure extracellular neural spikes from an animal with a neural amplifier and an analog-to-digital converter (ADC), or alternatively be directly injected with digitized pre-recorded neural voltages for system testing. The desktop computer contains three sub-processing units– 1) raw data smoothing, spike detection and feature extraction, 2) spike sorting using SPC and 3) template estimation. The FPGA module also contains four sub-processing units– 1) raw data smoothing, peak detection and spike isolation, 2) feature extraction, 3) neural spike sorting based on template matching, and 4) calculation of spike count statistics.
Fig 2.
A block diagram illustrating the hardware implementation of the spike detection and isolation.
An 8-sample smoothing filter was used to remove high frequency noise from the input neuron signal, followed by a peak detection module based on the NEO algorithm to detect a neural spike for isolation. A 64-sample FIFO was used to temporarily store the isolated data stream. A peak index counter and a peak height register worked synergistically to determine the peak index to correctly isolate the neural spike maximum. A 32-sample neural spike arrays centering against the spike peak center were outputted from the module for downstream feature extraction.
Fig 3.
Hardware implementation of the CM and ED classifiers.
Investigators can select one of the two classifiers through the “Select CM/ED” pin. Within the CM classifier, there are in total 8 covariance units (Cova) and 7 operator units (OperatorM) for determining the maximum correlation coefficient for the incoming spike to the eight cluster templates. Based on this design, the covariance calculations are performed in parallel to achieve minimum calculation latency. The hardware implementation of the covariance units, the operator units, and the ED classifier are also shown in detail on the top two sections and the bottom section of the figure.
Fig 4.
Estimate of maximum spike sorting rate and sorting latency for the system.
Two neural spikes extracted from the neural recording of a gerbil were pieced together with a time difference tdiff to create an artificial voltage trace, which was sent to the FPGA hardware to estimate the maximum spike sorting rate. As the time difference tdiff between two spikes was monotonically reduced to tmin, the FPGA hardware could no longer separate the two neural spikes and the voltage trace was considered as a single spike, resulting in missing classification for the second spike. The spike sorting latency of the system was also estimated by measuring between the time when the neural spike entered the FPGA for sorting and the time the FPGA resulted in a classification label for the neural spike.
Table 1.
Sorting latency of the FPGA based real-time spike soring module.
Fig 5.
Comparison of the spike sorting accuracies for CM (dark) and ED (white) under various noise contamination conditions.
Third party pre-labeled neural spikes were used to estimate the spike sorting accuracy of our system [42]. The first 20 sets of spike data were contaminated by Gaussian noise and were separated into four different groups–two groups (EasyGroup1 and EasyGroup2) constructed by spikes that are easily separable and the other two groups (Diff.Group1 and Diff.Group2) constructed by spikes with similar temporal profiles. The final three sets of test data were non-Gaussian noise contamination and were constructed to mimic spike shape changes caused by various physiological conditions (electrical drifting, cell bursting activity, and local field potential occurrence). The results indicate that CM can achieve higher sorting accuracies over ED, especially for neural spikes contaminated with non-Gaussian noise.
Fig 6.
Comparison between CM and ED using neural spikes artificially constructed to simulate electrode drifting.
(A) and (B) Temporal profiles of two artificial neural spike clusters with linearly decreased spike amplitudes. (C) Temporal profiles of the third artificial neural spike cluster with linearly increased spike amplitudes. (D) Sorting accuracies of the three neural clusters using CM and ED. (E) to (G) Correlational plots between each of the three correlational coefficients in CM and the figures show clean separation among the cluster groups along the diagonal line. (H) to (J) The diagonal line cannot separate the clusters in ED sorting and significant overlaps can occur.
Fig 7.
Real-time spike sorting results based on pre-recorded neural spikes from an anesthetized gerbil.
(A) 0.5 s of neural voltage trace recorded from the brain stem of an anesthetized gerbil. The green stars and red triangles at the top of the figure indicate the locations of neural spikes of two neurons co-recorded by the same electrode. (B) Temporal profiles of the two cluster templates of the two neurons estimated by SPC. (C) Phase plot of the two cluster groups (green star and red triangle) with each marker representing a neural spike. (D) The firing rates of the two clusters calculated over the 100 seconds of neural data by the FPGA hardware.
Fig 8.
Real-time spike sorting results based on an awake behaving mouse.
(A) Real-time spike template matching (ED and CM) compared to off-line k-means classification recorded from the olfactory bulb of an awake behaving mouse. The sorting agreement is higher than 80% for all clusters. (B) Temporal spike profiles (Trial 3) of four clusters sorted by off-line K-means, real-time ED and real-time CM. The numbers at the top of each plot indicate the number of spikes classified to the cluster, and the results indicate similar performance of the three techniques. The red arrow indicates a spike sorting anomaly, likely caused by using vector distances as the sole classification criterion in ED.
Fig 9.
Sorting accuracy and time comparison between CM and ED hardware template matching to other off-line spike sorting algorithms.
(A) Percentage ratios of correctly classified (black), misclassified (red) and unclassified (blue) neural spikes comparing several off-line neural spike sorting algorithms to CM and ED hardware template matching using the third party labeled neural spikes [42]. The results indicate that CM and ED achieve comparable sorting accuracies with the other off-line sorting algorithms. (B) Sorting time comparing between hardware CM and ED template matching to other off-line spike sorting algorithms. While CM and ED template matching requires less than 2 ms of sorting time, the off-line spike sorting techniques require sorting time in seconds. Other offline sorting methods used in the comparison: Phy [12,23], Wave_Clus [42], Bayes, SVM (Support Vector Machine), K-means, Artificial Neural Network (ANN) [60].
Table 2.
Performance summary and comparison with others work.