Mapping mesoscale axonal projections in the mouse brain using a 3D convolutional network

Abstract

The projection targets of a neuronal population are a key feature of its anatomical characteristics. Historically, tissue sectioning, confocal microscopy, and manual scoring of specific regions of interest have been used to generate coarse summaries of mesoscale projectomes. We present here TrailMap, a three-dimensional (3D) convolutional network for extracting axonal projections from intact cleared mouse brains imaged by light-sheet microscopy. TrailMap allows region-based quantification of total axon content in large and complex 3D structures after registration to a standard reference atlas. The identification of axonal structures as thin as one voxel benefits from data augmentation but also requires a loss function that tolerates errors in annotation. A network trained with volumes of serotonergic axons in all major brain regions can be generalized to map and quantify axons from thalamocortical, deep cerebellar, and cortical projection neurons, validating transfer learning as a tool to adapt the model to novel categories of axonal morphology. Speed of training, ease of use, and accuracy improve over existing tools without a need for specialized computing hardware. Given the recent emphasis on genetically and functionally defining cell types in neural circuit analysis, TrailMap will facilitate automated extraction and quantification of axons from these specific cell types at the scale of the entire mouse brain, an essential component of deciphering their connectivity.

Keywords: axons; light-sheet microscopy; neural networks; tissue clearing; whole-brain.

Fig. 1.

Overview of the TrailMap workflow to extract axonal projections from volumetric data. (A) Annotation strategy for a single subvolume (120 × 120 × 101 voxels). Three planes are labeled with separate hand-drawn annotations for background, artifacts, and axons. The one-pixel-width “edges” label is automatically generated. (B) Basic architecture of the encoding and synthesis pathways of the 3D convolutional U-Net. Microscope volumetric data enters the network on the left and undergoes a series of convolutions and scaling steps, generating the feature maps shown in gray. Skip connections provide higher-frequency information in the synthesis path, with concatenated features in white. More details are provided in ref. . (C) A network output thinning strategy produces skeletons faithful to the raw data but with grayscale intensity reflecting probability rather than signal intensity or axon thickness. XY and XZ projections of one subvolume are shown (122 × 609 × 122 µm). (D) A 2-mm-thick volumetric coronal slab, before and after the TrailMap procedure, which includes axon extraction, skeletonization, and alignment to the Allen Brain Atlas Common Coordinate Framework.

Fig. 2.

Comparison of TrailMap to a random forest classifier. (A) From left to right: 60-µm Z-projection of the probability map output of an Ilastik classifier trained on the same 2D slices as the TrailMap network; typical segmentation strategy, at a P > 0.5 cutoff; raw data for comparison; skeletonized axonal armature extracted by TrailMap; and probability map output of the TrailMap network. To better indicate where P > 0.5, color maps for images are grayscale below this threshold and colorized above it. (Scale bar, 100 µm.) Second row shows the same region as above, rotated to a X-projection. (B) Sparse axon identification by TrailMap and Ilastik; images show 300 µm of Z- or X-projection. (Scale bar, 100 µm.) (C) A 3D maximum intensity projection of a raw volume of axons, the resultant Ilastik probabilities, and TrailMap skeletonization. (Scale bar, 40 μm.) (D) Network output from examples of contaminating artifacts from nonspecific antibody labeling, similar to those included in the training set. Raw data, Ilastik segmentation, and TrailMap skeletonization are shown. (Scale bar, 200 μm.)

Fig. 3.

Volumetric visualizations highlight patterns of axonal innervation. (A) Coronal Z-projection of extracted serotonergic axons, color-coded by depth (0–500 µm) and overlaid on the CCF-aligned serial two-photon reference atlas. SI Appendix, Fig. S1 details viral-transgenic labeling strategy. (B) The same volumetric slab as in A presented as a density heatmap calculated by averaging a rolling sphere (radius = 225 μm). Green arrow highlights a density hotspot in the amygdala. (C) TrailMap-extracted serotonergic axons innervating forebrain are subdivided and color-coded based on their presence in ABA-defined target regions. (D) Same brain as in C as seen from a dorsal viewpoint, with major subdivisions spatially separated. Midline is represented by a dashed white line. Inset highlights the region indicated by the dashed box in A; Z-projection, 500 μm. (Scale bar, 200 μm.) (E, Left) Mesh shell armature of the combined structures of the lateral (LA), basolateral (BLA), and basomedial (BMA) amygdala in coronal and sagittal views. (Right) Density heatmap and extracted axons of the amygdala for the same views. LA, BLA, and BMA are color-coded in shades of green/blue by structure, and axonal entry/exit points to the amygdala are colored in red. White arrows highlight the same density hotspot indicated in B.

Fig. 4.

Generalization to other cell types with and without transfer learning. (A) Dorsal view of posterior brain with extracted axon collaterals from pons-projecting cerebellar nuclei neurons color-coded by their presence in major subregions. Injection sites in the right hemisphere lateral, interposed, and medial CN from the same brain are indicated. (B) Coronal (Top) and sagittal (Bottom) views of the extracted axons (orange) within the structure of thalamus (cyan) from the brain in A. (Scale bar, 500 μm.) Zoomed images (Right) of midline thalamus show exclusion of axons from specific subregions. VPPC, ventral posterior nucleus, parvicellular part; VPM, ventral posteromedial nucleus; Po, posterior thalamic nuclear group. Midline shown as dashed vertical line. (Scale bar, 200 μm.) (C) Extracted thalamocortical axons in barrel cortex. XZ-projection of a 3D volume extracted from a flat-mount imaged cortex. Dashed lines indicate upper and lower bounds of cortex. Z-projection, 60 μm. (Scale bar, 100 µm.) (D, Left) Raw image of axons of prefrontal cortex neurons in posterior cortex. VIS, visual cortex; ECT, ectorhinal; ENTl, lateral entorhinal. (Right) Axons extracted by TrailMap model before (green) and after (red) transfer learning. Z-projection, 80 μm. (Scale bar, 100 μm.) Labeling procedure, virus, and transgenic mouse for all panels are described in SI Appendix, Fig. S1 and Methods.