Extended Data Fig. 5: Deep-learning feature detection and control point motion correction.
From: Long-term optical imaging of the spinal cord in awake behaving mice
a, Comparison of reference frame 42 (cyan) to movement frame 804 (red, overlaid on cyan image) before and after LD-MCM motion correction. Scale bar, 300 µm. b, Example of DLC-identified vascular features used for cross-session registration (DLC model trained using day 41). Scale bar, 300 µm. c, Model error as a function of DLC iterations (500,000 iterations, n = 4 mice). d, Spearman’s correlation of each feature to other features in a movie from a Phox2a-Cre; Ai162 mouse. Green arrow, a feature that has reduced correlation with all other features and can thus be removed to improve motion correction. e, Point clouds with each dot (2001 frames) represents the rostrocaudal and mediolateral location of that feature on an individual frame during an imaging session (~6 min, 13.9 Hz, mouse from a). f, DLC tracks (1) large mediolateral shifts in the field of view (yellow arrow) and (2) camera errors that result in a split of the field of view (yellow line). Only showing features with confidence >0.1. Scale bar, 300 µm. g, Labeling (DeepLabCut, 20 frames from day 75) of vascular features in a Phox2a-Cre; Ai162 (GCaMP6s) mouse across 52 neural activity imaging sessions, spanning nearly 5 months. Scale bar, 300 µm. h, Feature locations (normalized to the session mean location) across 13 features tracked in raw and LD-MCM motion corrected movies.Green lines, frames shown in i. i, Frames before and after LD-MCM motion correction. Yellow dots: tracked features with the line showing connected features indicating improvement with LD-MCM. Scale bar, 300 µm. j, Performance of LD-MCM as a function of the number of features used for control point registration (n = 10 movies, n = 2 mice). Mean, median, and standard deviation calculated per movie for each combination of imaging session, parameter value, trial, and feature. Then the mean is taken across all features for the final displayed values (each data point). Boxplots in all figures display the 1st, 2nd (median), and 3rd quartiles with whiskers indicating 1.5*IQR; outliers are omitted.
