Figures and figure supplements Computer assisted detection of axonal bouton structural plasticity in in vivo time-lapse images

The ability to measure minute structural changes in neural circuits is essential for long-term in vivo imaging studies. Here, we propose a methodology for detection and measurement of structural changes in axonal boutons imaged with time-lapse two-photon laser scanning microscopy (2PLSM). Correlative 2PLSM and 3D electron microscopy (EM) analysis, performed in mouse barrel cortex, showed that the proposed method has low fractions of false positive/negative bouton detections (2/0 out of 18), and that 2PLSM-based bouton weights are correlated with their volumes measured in EM (r = 0.93). Next, the method was applied to a set of axons imaged in quick succession to characterize measurement uncertainty. The results were used to construct a statistical model in which bouton addition, elimination, and size changes are described probabilistically, rather than being treated as deterministic events. Finally, we demonstrate that the model can be used to quantify significant structural changes in boutons in long-term imaging experiments.

reconstructions and 2PLSM maximum intensity projections for the four axons included in CLEM analyses (same as shown in Figure 4).Rows 2-8 show intensity profiles of these axons generated by using raw voxel intensities along the trace, mean filters of box sizes 3 Â 3 Â 3 and 5 Â 5 Â 5, Gaussian filters [see Equation ( 1)] of sizes R = 1 and R = 2, and median filters of box sizes 3 Â 3 Â 3 and 5 Â 5 Â 5.Each profile was normalized by its median profile intensity.Row 9 shows multi-scale LoG xy profiles normalized by shaft intensities.In the absence of filtering, profiles appear jagged, which may lead to false positives in peak detection (arrows).On the other hand, large filters (5 Â 5 Â 5 and R = 2) may not resolve small boutons or closely positioned boutons (arrows).When the filter size is carefully tuned (3 Â 3 Â 3 and R = 1 in this case) profiles generated by various filters may appear similar to those obtained with the multi-scale LoG xy .(E) Bouton measurements based on these profiles were compared to bouton volumes measured in EM.LoG xy filter profile normalized with shaft intensity leads to both accurate bouton detection and highest degree of correlation between peak amplitudes on the profiles and bouton volumes.

Figure 1 .Figure 2 .Figure 3 .
Figure 1.Challenges in LM-based bouton detection and measurement.(A) Maximum intensity xy projection of an image stack showing axons of fluorescently labeled neurons in superficial layers of mouse barrel cortex.High density of labeled axons makes it difficult to automatically detect boutons and track their structural changes over time.Scale bar is 20 mm.(B) A subset of labeled axons from the region outlined in (A).To improve visibility, image intensity beyond five voxels from the axon centerlines was set to zero.Bouton detection and bouton size measurement are confounded by large variations in fluorescence levels across axons.(C) Axons from (B) shown on the zx maximum intensity projection.Horizontal scale bars in (B) and (C) are 5 mm.Vertical scale bar in (C) is 15 mm.Lower resolution in z compared to xy is yet another challenge in bouton analyses.(D-F) Magnified views of the highlighted boutons from (B).Close proximity of boutons on an axon (D), large range of bouton sizes (E), and large range of bouton fluorescence levels (D-F), present additional obstacles to accurate bouton detection and measurement.Scale bar in (D-F) is 1.25 mm.DOI: https://doi.org/10.7554/eLife.29315.002

Figure 5 .
Figure5.Probabilistic definition of an LM bouton based on measurement uncertainty derived from short-term imaging experiments.The same set of axons was imaged 7 times within 80 min with various microscope settings and cranial window conditions (inset in A).Putative boutons detected based on the first imaging session (condition A) were chosen to be the gold standard.Precision (A) and recall (B) in bouton detection were measured under the remaining conditions, B-G.Both precision and recall increase with bouton weight.While for very small boutons (w < 2.0, dashed line) detection is unreliable, agreement with the gold standard is achieved across all imaging conditions in 95% of boutons with weights greater than 2.0.Numbers of boutons in the gold standard are indicated next to the data points in (A).Inset in (B) shows an example of one axon segment imaged in conditions A-G. (C) Bouton weights under different imaging conditions are plotted against the gold standard weight.Best fit lines show no significant bias for conditions B and C, however small, but significant reduction in mean bouton weight was observed in the remaining four conditions (all p < 0.03, twosample t-test).Abbreviations used in the inset of A: LP is laser power in mW, PMT denotes photomultiplier tube voltage in Volts, and WC is cranial window condition, where 'n' stands for normal and 'a' indicates presence of a thin layer of agarose.Color code used in (A-C) is defined by the inset tablein (A).(D) CDFs for differences in bouton weights across imaging conditions.Data from all conditions were pooled.Different lines show CDFs for various intervals of mean bouton weight.(E) Variance in bouton weight difference increases linearly with mean bouton weight ( 2 linear regression with var Dw ð Þ ¼ a w h i, p = 0.33, a = 0.24 ± 0.01, mean ± s.d.).Error-bars indicate standard deviations obtained with bootstrap sampling with replacement.(F) Red line shows the distribution of true bouton weight for a putative bouton of measured weight w = 1.5.Area under the curve to the right of w threshold = 2.0 gives P boutonjw ð Þ = 0.12.Large putative boutons (e.g.blue curve, w = 3.0) have high probability of being LM boutons.DOI: https://doi.org/10.7554/eLife.29315.011

Figure 6 -
Figure 6-figure supplement 1. Matching boutons in time-lapse images.(A).xy projection of an axon segment that was imaged at 4 day intervals over a period of 68 days.Image intensity in each session is normalized independently as described for intensity profiles.Colored lines link corresponding putative boutons.These lines do not always run in parallel due to small rotations of the brain, nonlinear tissue distortions, and movements of boutons.(B).Normalized and aligned intensity profiles of the axon from A (see Materials and methods for details).Figure 6-figure supplement 1 continued on next page

Figure 6 -
Figure 6-figure supplement 1. Matching boutons in time-lapse images.(A).xy projection of an axon segment that was imaged at 4 day intervals over a period of 68 days.Image intensity in each session is normalized independently as described for intensity profiles.Colored lines link corresponding putative boutons.These lines do not always run in parallel due to small rotations of the brain, nonlinear tissue distortions, and movements of boutons.(B).Normalized and aligned intensity profiles of the axon from A (see Materials and methods for details).Figure 6-figure supplement 1 continued on next page