With around 100 million cells in the mouse brain and sizes of whole-brain cellular resolution images approaching the terabyte scale, advanced image analysis tools are needed to accurately quantify cells. Our computational pipeline can preprocess images and quantify nuclei within the mouse cortex while maintaining a reasonable compromise between cell-detection accuracy, imaging time, and computational resources. Demonstrating the procedure will be Felix Kyere and Ian Curtin, graduate students from my lab.
To begin, mount the sample in the correct sample size holder such that the sample is oriented with the z-dimension no more than 5.2 millimeters in depth due to the rated working distance of the UltraMicroscope II microscope. Then insert the holder into the sample cradle such that the screw of the holder is at 45 degrees angle to the supports of the cradle. Next, place the cradle into a position such that the sample is oriented perpendicular to the light path.
Afterward, set the zoom body on the microscope to 4x magnification or higher, yielding 0.75 micrometers per pixel. In the Inspector Pro software, select a single light sheet with a numerical aperture value of approximately 0.08. To ensure axial resolution is maintained along the width of the image, select Horizontal Dynamic Focusing and apply the recommended number of steps depending on the laser wavelength.
Then adjust the fine focus for each channel with respect to the registration channel and laser power per channel with respect to the channel properties. Next, adjust the light-sheet width to about 50%to ensure sheet power is distributed optimally in the y-dimension for the sample size. Afterward, set the number of tiles with respect to the size of the sample with a recommended overlap of 15%between tiles, and capture images for each channel sequentially for each stack at a given tile position.
First, download and install Conda environment manager for Linux and NuMorph image-processing tools. On the command line, run matlab and NM_setup. m from NuMorph to download and install image analysis software packages needed for analysis.
Then specify sample names, input and output directories, channel information, and light-sheet imaging parameters by editing the file NM_samples.m. For intensity adjustment, in NMp_template, set intense the adjustment to true and use_processed_images as false when working with a new set of images. Next, set save_images and save_samples as true.
Next, set tile shading to basic to apply shading correction using the basic algorithm, or manual to apply tile shading correction using measurements from the UltraMicroscope II at specific light-sheet widths. For image channel alignment, in NMp_template, set channel_alignment to true and channel alignment_method to either translation or elastics. Next, set sift_refinement as true and overlap value of 0.15 to match tile overlaps during imaging.
To run preprocessing steps in MATLAB, specify the sample name and set the configuration to NM_config process sample. Then run any of the preprocessing steps by specifying NM_process config stitch while specifying the stage using intensity, align, or stitch, and check the output directory for output files for each of the stages. Start with a 3D Atlas image and an associated annotation image that assigns each voxel to a particular structure.
Align both the Atlas image and annotation file to ensure they match correctly in the right orientation. After alignment, process the files in NuMorph to specify the inputs as described in the manuscript by executing the command. In NMa_template, set resample_images to true and resample_resolution to match the Atlas.
Then specify the channel number to be resampled using resample_channels. Afterward, set register_images to true, specify the atlas_file to match the file in the Atlas directory, and set registration_parameters as default. Then set save_registered_images to true.
For nuclei detection, cell counting, and classification, set both count_nuclei and classify_cells as true. Then set the count_method to 3dunet and min_intensity to define a minimum intensity threshold for detected objects. Next, set classify_method to either threshold which is based on a non-supervised fluorescence intensity at centroid positions, or SVM which models a supervised linear support vector machine classifier.
To perform analysis steps in MATLAB, specify the sample name and set the configuration to NM_config analyze sample. Next, run any of the analysis steps by specifying NM_analyze config stage while specifying the stage using resample, register, count, or classify and check the output directory for output files for each of the stages. In NMe_template, set update to true and compare_structures_by to either index.
Then set the template file and structure table which specifies all possible structure indexes and structures to evaluate while specifying cell counting and cell type classification. Tissue clearing using iDISCO+protocol and neuronal layer-specific nuclei markers resulted in clearly-defined cell groups of upper and lower-layer neurons in the isocortex. Cell counting using NuMorph was dependent on successful preprocessing steps involving intensity adjustment, channel alignment, and stitching.
However, errors in preprocessing steps could result in improper stitching, leading to improper alignment and stitching, and thus, result in images with in-focus and out-of-focus pattern. In order to count nuclei from specific brain regions, the stitched images will be annotated using Atlas, allowing annotations to be overlaid on brain regions. The centroids of nuclei were detected with a trained 3D U-Net model in NuMorph with around 12 million total nuclei that were TO-PRO-3-positive in the isocortex with about 2.6 million brain-two-positive and 1.6 million CTIP2-positive nuclei.
About 3.7 and 2.9 million TO-PRO-3-positive total nuclei in the basal ganglia and hippocampal allocortex were detected, respectively. However, the brain-two-positive cells detected in these two brain regions were negligible, and only about 1.5 in less than one million CTIP2-positive cells were detected each in the basal ganglia and hippocampal allocortex. Perform visual quality checks during acquisition to achieve good segmentation.
And properly set up the Conda environment to ensure that no errors are encountered in a downstream analysis. In addition to cell counting, this pipeline allow for integration with other segmentation tools that measure cell size and shape which can then be compared across genotype groups. With our pipeline, we can identify how brain anatomy changes at cellular resolution, leading to identification of cell types and brain regions important for disease risk.
