Neuronal morphometry directly from bitmap images

TLDR: The use of the Sholl Analysis software to quantify arborization directly from bitmap images correctly identified 80–86% of cells, and the utility of the method in tackling neurons that are particularly slow to reconstruct manually was explored.

Abstract: To the Editor: Neuroscientists measure the tree-like structures of neurons in order to better understand how neural circuits are constructed and how neural information is processed. In 1953, Donald Sholl published his well-known technique for quantitative analysis of the complex arbors of dendrites and axons1, but conventional methods still require reconstruction of arbors via time-consuming manual or semi-automated tracing from microscopy images. To bypass this reconstruction step and perform the Sholl technique directly on images instead, we developed Sholl Analysis (http://fiji.sc/Sholl), an open-source program for ImageJ/Fiji2 (Supplementary Fig. 1). The plug-in employs an improved algorithm to retrieve data from twoor three-dimensional (2D or 3D) bitmap images in any format supported by the Bio-Formats library (Supplementary Methods). It pairs this data retrieval with curve-fitting, regression analysis and statistical inference so that users can automatically extract a collection of Sholl-based metrics of arborization1,3 (Supplementary Note). Using individual cortical pyramidal neurons in 3D images, we found Sholl Analysis to be accurate when benchmarked against corresponding manual reconstructions (Supplementary Fig. 2). The method was also resilient to image degradation by simulated shot noise (Supplementary Fig. 3 and Supplementary Software). To further assess accuracy, and to explore the utility of Sholl Analysis in tackling neurons that are particularly slow to reconstruct manually, we studied cerebellar Purkinje cells in mice, which have large and intricate dendritic arbors. From tiled 3D image stacks of cerebellum (Fig. 1a), we selected seven Brainbow2.1-expressing Purkinje neurons and isolated their morphologies (Fig. 1b and Supplementary Note). We then used the Sholl Analysis software to retrieve ten metrics and found they were indistinguishable from those retrieved from manual reconstructions of the same 7 cells (Fig. 1c,d and Supplementary Methods). To probe the sensitivity of the Sholl Analysis software, we asked whether its metrics could be used to distinguish closelyrelated neocortical interneuron subtypes. Parvalbumin-positive (PV) interneurons in layer 5 of visual cortex can be morphologically classified into two subtypes on the basis of their axonal morphology: type 1 PV cells have ascending axons arborizing in layer 2/3, whereas axons of type 2 cells remain in layer 5 (ref. 4). Because their dendritic arbors are indistinguishable4, these two cell types otherwise appear highly similar (Fig. 1e,f). Using the Sholl Analysis software, we retrieved 18 metrics directly from 3D image stacks of 12 PV interneurons. We then used Ward’s hierarchical clustering based on these metrics to independently classify these cells (Fig. 1g and Supplementary Fig. 4). The 12 cells segregated into two groups: one group of five neurons and another of seven. We found that all the neurons but two were correctly classified, with one cell assigned incorrectly to each class (Fig. 1g). Thus, our use of the Sholl Analysis software to quantify arborization directly from bitmap images correctly identified 80–86% of cells. In agreement, linear Sholl plots of type 1 cells indicated more branching than was found for type 2 cells at a distance of 225–300 μm from the soma (Fig. 1h), which corresponds to check and inviting routine use. Second, the software can generate a summary report of the current system performance or a full report containing all individual PSF measurements and associated fitting parameters. Third, a table with the extracted resolution, planarity and colocalization data can be exported. This can be used for subsequent analysis, such as in an image processing or restoration pipeline. In addition, an average PSF from a user-selectable region of interest can be exported, for example, for image deconvolution. We used PSFj to quantify the performance of various high– numerical aperture (NA) objectives and to track day-to-day and system-to-system variation. The results showed substantial performance differences and allowed us to identify strengths and weaknesses of individual objectives as well as general shortcomings (Supplementary Figs. 1 and 2). In particular, we found that whereas lateral resolution performance generally fell short (~20–30%), axial resolution often met or exceeded expectations from the scalar approximation of the PSF commonly used in textbooks2 (Supplementary Note). Planarity was usually well corrected with variations over the FOV below the axial resolution and allowed for the detection of tilted slides caused, for example, by dust particles or misaligned slide holders or stages. Axial chromatic shifts were usually small, with little variation across the FOV (Supplementary Table 1). In contrast, chromatic shifts often showed circular symmetry and increased toward the edge of the FOV, which is a sign of lateral chromatic aberrations. Day-to-day performance variation of most objectives was relatively small (~2–6%) and comparable to single-measurement FOV variations (Supplementary Table 2). Furthermore, testing a limited number of identical objectives identified objective-toobjective and microscope-to-microscope variations of about 10% (Supplementary Tables 3 and 4). The PSFj software is open source and based on libraries from various sources, including ImageJ3 and μManager4, and it runs as a stand-alone application on the three major operating systems (using Java).