Cintiq 21ux

Segmentation was performed on filtered tomograms with the default parameters of EMAN2 (low-pass gaussian cutoff of 0.25 and high-pass gaussian cutoff of 5px) and Convolutional Neural Networks (CNN)³¹ (link) were used to recognize the fibers in the tomograms (Figures S1A–S1C). Training was performed on several tomograms by boxing ∼20 positive examples and ∼100 negative examples. The positive examples were precisely segmented using a graphical tablet (Wacom Cintiq 21uX) and the CNNs were trained with the default parameters except for the learn rate that was increased in some instances to 0.001 instead of the default 0.0001. The outcome of the trained CNN was checked on the boxed particles and if satisfactory the CNN was applied on the tomogram. Eventually, a second round of training was performed with additional boxes from another tomogram from the same dataset or on itself. The resulting CNN map was then carefully examined versus the filtered tomogram to ensure they agreed, and segmentation was specific to the fibers. For tomograms acquired over the same session on the same lamellae, the same CNN was able to generalize well and segment accurately. Tomograms from different datasets and different lamellae usually required retraining a CNN.
Satisfactory CNN segmented volumes were then transferred into Amira (Thermo Fisher) to perform template matching fiber tracing with the TraceX Amira plugin³² (link) (Figures S1D–S1F) in order to model the fibers as a set of connected nodes. To be able to optimize parameters, we reduced the processing time by binning twice (binning 8 total) the CNN maps. The first step, Cylinder Correlation, was performed with the following starting parameters: cylinder length of 50 pixels, an angular sampling of 5, and missing wedge compensation was toggled. The diameter of the template (outer cylinder radius) was set to closely match the apparent diameter of the fibers in the tomogram, usually 4 pixels. As advised in the Amira user guide section 3.8 on the XTracing Extension, the mask cylinder radius was set to 125% of the outer cylinder radius. The outcome was visually checked to see if the fibers were detected correctly and not too many artefacts were generated. Parameters were slightly modified one-by-one if needed to improve the output. The subsequent step, Trace Correlation Lines was performed with the following nominal parameters: minimal line length 60 pixels, direction coefficient 0.3, and minimal distance of 2-times outer-cylinder diameter used previously. Minimum seed correlation and minimum correlation are tomogram-dependent parameters. These values were defined on the correlation field by defining the reasonable correlation value range. The minimum seed correlation and minimum continuation quality are the upper and lower limits of the range, respectively. For the search cone, length was set to 80, angle to 37°, and minimal step size was 10%. The outcome was visually checked to see if the fibers were being traced correctly. To do so, we used the Spatial Graph View function and checked for artificial fiber trackings. Parameters were modified if needed to enhance fiber detection and reduce false discoveries. Because of the inherent nature of the signal of cryo-ET volumes and their CNN maps, punctate signals would generate and propagate artefactual vertical (parallel to the Z-dimension) lines. These were first selected by using a Tensor XZ and Tensor ZZ visualizer in the Spatial Graph View window and identifying the appropriate thresholds. After the coordinates of all fibers were extracted as a.xml file, fiber tracks with values above/below the thresholds were trimmed out.

To reconstruct the course of V1 and its subbranches, a series of beak slices stained against CtB were systematically aligned and normalized based on hand drawings in a coordinate grid using a pen display (three animals; Wacom Cintiq 21UX, Wacom, Düsseldorf, Germany) and the Adobe Illustrator 25.4.1 software (Adobe Systems Software, Dublin, Ireland, RRID: SCR_010279). The course of V1 and its subbranches was translated into 2D reconstructions, one depicting the course of the subbranches as a side view (Figure 1A) showing the dorsoventral coordinates of the nerve subbranches, and one as viewed from above (Figure 1B) to show the mediolateral coordinates. Anatomical boundaries within the trigeminal brainstem complex were defined based on previous studies on the known restricted expression of the immediate early gene Egr-1 and on morphometric features using the general neuronal marker HuC/HuD (Heyers et al., 2010 (link); Lefeldt et al., 2014 (link); Elbers et al., 2017 (link); Faunes and Wild, 2017b (link); Kobylkov et al., 2020 (link)).