Precision in 3D: A Fast and Accurate Algorithm for Reproducible Motoneuron Structure and Protein Expression Analysis

Figures

• 

  Figure 8. Threshold sensitivity of macro-cluster measures for user 1. (A,B) Net expression. (C,D) Relative net expression. (E,F) Cluster Size. User 2 threshold sensitivity graphs looked nearly identical to user 1 and had no notable differences, so we chose to only display user 1’s results for clarity and simplicity. Each point represents the median of each animal at the indicated threshold shift. Gray shaded regions are the average steady threshold range for all cells, the midpoint of which was used for the auto-threshold applied throughout the results (each cell had its own range and midpoint, with midpoints ranging from 12.5% to 27.5%). Gray arrows indicate the average percent change in the parameter across the steady threshold range, and black arrows indicate the average percent change across the entire (10–30%) threshold range.

• 

  Figure 9. Three-dimensional visualizations at some key threshold shifts. One example cell per animal (each row) is displayed with algorithm-identified clusters at each relevant threshold shift (each column). At the 5% shift, we see over-expression and large clumps of clusters. At the 40% shift, we see bare patches of soma membrane and under-represented macro-clusters. Our 10% default shift appears reasonable but can lead to over-labeling of some cells (one example is the displayed cell of animal A). We found our auto-threshold shift for optimum cluster measure stability better handles these difficult-to-threshold cells without changing the trends of our data. Our auto-threshold enables our hypothesis statistics to be less manipulated by over-labeling that can unexpectedly occur on portions of some membranes due to random label artifact or other image conditions that are both tedious and difficult to identify visually and can be inconsistent from one user’s reconstruction to the next.

• 

  Figure 1. Algorithm box diagram. The image files on the left and user inputs on the right are fed into the central algorithm. Parameters associated with the 3D soma reconstructions are indicated by blue backgrounds, and parameters associated with the protein clusters are indicated by green backgrounds. The algorithm quantitates the auto-identified soma and protein cluster structures and then outputs the soma measures and protein cluster measures. These outputs can be compared across multiple combinations of user inputs by simply looping the algorithm through different scenarios. All data is organized and saved to Microsoft Excel workbooks for easy plotting and hypothesis testing.

• 

  Figure 2. Visualization of one cell and sampling comparisons between manual and algorithm methods. Manual methods with existing software, showing (A) one manually traced blue soma (trace 11, measuring the 2D largest cross-ectional area (LCA)) and 8 smaller manually traced en-fosse white C-bouton protein clusters (traces 3–10) in Fluoview; and (B) 3D rendering of the blue soma and white C-boutons in Neurolucida. (C) Stack of our 2D auto-generated soma outline (red ROIs) and identified C-bouton protein label displayed in green. (D) Our 3D visualizer with auto-generated soma reconstruction and auto-identified protein clusters. The red portion in the middle of the cell indicates the slices on which manual cluster counts would be measured to approximate cluster density. (E) The average number of clusters measured per cell for cluster size comparisons in manual vs. algorithm methods. Manual average was taken from Dukkipati et al., 2016 study [2] from 38 control cells. Algorithm average was from 60 control cells (average of two users’ measurements for each cell).

• 

  Figure 3. Evaluating algorithm accuracy across different geometric shapes. LCA, surface area, and volume of known 3D shapes of 3 different sizes (small, medium, and large). The left axes indicate the sizes, graphed as lines, with black for mathematically calculated, and red or blue for 2D or 3D algorithm measurements, respectively. The right axes indicate percent accuracy of the measurements compared to the calculated sizes, displayed as gray bars. Purple overlay indicates the algorithm measured larger (overestimation), while the orange striped filling indicates the algorithm measured smaller (underestimation). The orange box emphasizes the consistent volume under estimations for the curved geometries. The bottom panel between the legends is a 2D visual example of a pixel–curvature error, which becomes a voxel–curvature error in 3D. White areas within the dashed circle add to underestimation, and gray areas outside the circle add to overestimation.

• 

  Figure 4. Algorithm comparison to manual measurements. (A) Algorithm soma LCA correlates to manual measurements (p < 0.0001) and measures 8.2% smaller than the manual traces (paired Wilcoxon, p < 0.0001). (B) Algorithm C-bouton cluster LCA correlates to manual measurements (p = 0.0017) and measures 72.4% smaller than the manual traces (paired Wilcoxon, p < 0.0001). (C) The 3D algorithm density using 100% of the soma membrane correlates to the manual density approximation from 25% of the membrane for the same 20 cells, but the strength of this correlation varies depending mostly on the manual user (notice that the linear regression lines form 2 groups of blue and pink). (D) Manual and algorithm densities compared pairwise for the same 20 cells between the 2 users. The manual density measurements vary between users by an average of 51.2%, while the algorithm density measurements vary between users by an average of 16.4% (p < 0.05 *, p < 0.01 **, and p < 0.001 ***).

• 

  Figure 5. Visualizing algorithm cluster LCAs vs. manual traces. (A) Manual traces of 4 example clusters (traces 6, 7, 5, and 3), with the manual LCA in white text under the cluster. Notice the tendency toward an elliptical shape in all (but especially the first 3) clusters. Also, notice some very dim pixels near the inside cluster edges, suggesting these clusters were traced with an increased brightness for visualization and are thus overestimated. There are also some dim pixels within the centers of the clusters, which would be included as part of the labeled area in the manual LCA. (B) The same clusters overlayed with bright green as identified and measured by the algorithm at the 10% default threshold shift. The red lines are nearby soma edges. In this visualizer, the original C-bouton label intensity is displayed in dark blue. Dark blue indicates pixels that had visual label but were not included as part of the cluster because they (1) were too far from the soma membrane, (2) were not bright enough to meet the threshold for label, or (3) did not meet the criteria to be a part of that cluster. The single dark blue pixel in these clusters (specifically in the second from the left…

• 

  Figure 6. Reproducibility and specificity of algorithm soma measurements. (A) LCA and (B) volume measured by two different users of the algorithm, analyzing their separate reconstructions of the same cells (p < 0.05 *, p < 0.01 **).

• 

  Figure 7. Reproducibility of algorithm macro-cluster measurements. All measures were gathered independently by 2 separate users, with their own cell reconstructions. (A) Number of clusters, (B) total cluster volume, and (C) cluster density (# clusters per soma membrane size) were less in animal A than animals B and C. (D) Relative total cluster volume (total cluster volume per soma membrane size) was not different between the 3 control animals. The cluster size ((E) 2D LCA and (F) 3D volume) was not different between the 3 control animals. (G) Soma membrane size was smaller in animal A than animals B and C, underlying part of the net expression differences in subfigures (A,B) (p < 0.05 *, p < 0.01 **, and p < 0.001 ***).