Fig 1.
Limitations of center of mass localization.
A: Visualization of localization algorithms for a point-source neuron equidistant to three electrodes located on a grid at (0,0), (1,2), and (2,1), where the electrodes will all receive the same signal strength from the neuron in ideal recording conditions (no noise or electrode degradation), given the neuron is located equidistant to all three electrodes. MT and GC are able to correctly estimate the “true” location while COM is not. B: Simulated electrophysiological data (MEArec), where green dots denote true neuron locations, red dots denote estimated neuron locations, and blue lines connect the true and estimated locations for a given neuron. Since COM is a weighted average of electrode locations, it forces all location estimates to be within the convex hull of the electrode array (all estimated neuron locations are within red square).
Fig 2.
SSL algorithm performance with no electrode degradation.
A: Visualization of template and spike localization estimates on simulated dataset (MEArec). The estimated spike locations are represented by translucent dot clouds, color-coded by their corresponding neuron. The estimated template location and true neuron locations are represented by solid red and green dots, respectively, and are not color-coded by neuron. The electrode array (shown excluding the “dead” electrodes to illustrate electrode degradation) is denoted by light gray dots. B: Visualization of template and spike localization estimates on experimental dataset (SPE-1). C: Performance metrics on simulated dataset, for both templates (left) and spikes (right). Performance is assessed as (i) percentage of estimates within 30µm of true locations (accuracy), (ii) speed of algorithm on all templates/spikes (runtime), and (iii) Euclidean distance between estimates and true location (median localization error and violin plot of individual localization errors). Bars represent mean metric across multiple simulations; error bars represent standard deviation of metric. Statistical significance was assessed using one-way ANOVA followed by Tukey’s HSD post-hoc tests; all metrics exhibited significant differences (ANOVA p < 0.05). D. Performance metrics on experimental dataset, for both templates (left) and spikes (right). Performance is assessed using accuracy and runtime; since we have lower resolution of true locations in experimental data, we use a higher tolerance for accuracy (deemed correct if estimate within 50µm of estimated true location). We additionally do not include runtime, as the experimental neurons were recorded across different recording sessions (we concatenate onto one probe in Fig 2B for presentation purposes), which does not allow for localization algorithm to amortize certain overhead costs across neurons. Bars represent metric across all experimental data.
Fig 3.
SSL algorithm robustness on simulated data (MEArec).
A: Template localization performance metrics at varying levels of electrode degradation. Performance is assessed as (i) percentage of estimated locations within 30µm of true locations (accuracy), (ii) median Euclidean distance between estimated and true locations (localization error), and (iii) estimated drift from original location. Estimated drift is the Euclidean distance between the estimated template location without degradation v. the estimated template location with degradation; i.e., a measure of how unstable the localization algorithm is against degradation. B: Spike localization performance metrics at varying levels of electrode degradation. Performance is assessed via accuracy and localization error. C: Visualization of template and spike localization estimates at varying levels of electrode degradation. Left panel is zoomed-in view of visualization outlined in red.
Fig 4.
SSL algorithm robustness on experimental data (SPE-1).
A: Template localization performance metrics at varying levels of electrode degradation (measured as the percentage of electrodes replaced with Gaussian noise). Performance is assessed as (i) percentage of estimated locations within 50µm of true locations (accuracy), (ii) median Euclidean distance between estimated and true locations (localization error), and (iii) estimated drift from original location. Since we have lower resolution actual location in experimental data v. simulated data, we use a higher tolerance for accuracy v. simulated data (50µm v. 30µm). B: Spike localization performance metrics at varying levels of electrode degradation. Performance is assessed via accuracy and localization error. C: Visualization of template and spike localization estimates at varying levels of electrode degradation. Left panel is zoomed-in view of visualization outlined in red.
Fig 5.
Impact of localization parameters on median spike localization error.
For MT, we examine the parameters “radius_um” (denoting the radius around the peak channel where signal is considered in optimization) and “max_distance_um” (denoting the max distance from the COM initial estimate considered as a solution to optimization). For GC, we examine the parameters “radius_um” (similarly denoting radius around the peak channel where signal is considered in optimization) and “percentile” (denoting percentage of the best scalar products used to estimate location). For COM, the parameters have less meaningful impact on algorithm performance and so we present the median spike localization error under default parameters. Median error using the default parameters are marked in red boxes, and we examine parameters an order of two lower and higher than the default parameters. In the first four columns, we show the progression of the median localization error as electrode degradation increases, and the colors denoting magnitude of error are normalized within the individual electrode degradation level but across the three localization methods (i.e., normalized within the “column”). We additionally show the change in localization error from 0% to 94% degradation for each method and parameter set in the fifth column, to indicate how much electrode degradation impacts performance. A: Results on the simulated dataset (MEArec). B: Results on the experimental dataset (SPE-1).