Organoid segmentation

Given images of nuclear staining in TIFF format, SCOUT contains tools for segmenting orgnaoid ventricles as well as the overall organoid shape (foreground).

Downsampling images

Since the segmentation of ventricles and whole-organoids is not a single-cell problem, the nuclear stain images are first downsampled to pixel width of 4 micron. For example, if the lateral dimensions of the original voxels are 0.67 micron, then a downsampling factor of 6x in x and y would give a 4 micron pixel resolution in each z-slice. To downsample the preprocessed TIFFs, the following command can be used:

scout segment downsample color0_rescaled/ syto_down6x 6 6 -v -t

This command will create a new folder called syto_down6x containing TIFF images downsampled 6x in y and 6x in x (the image dimension order is (y, x) for 2D and (z, y, x) for 3D). The flag -t specifies that the input folder contains a series of 2D TIFFs. Next, the downsampled 2D TIFFs can be stacked into a single 3D TIFF:

scout segment stack syto_down6x/ syto_down6x.tif -v

This downsampled 3D TIFF of the nuclear staining is the input for ventricle and whole-organoid segmentation.

Ventricle segmentation

SCOUT uses a pre-trained U-Net model to segment ventricles from nuclear stain images at 4 micron pixel resolution. Note that tensorflow must be installed in the scout conda environment for this to work, and the GPU-accelerated version will greatly improve performance. If the organoid images are larger than 1024 x 1024, then a GPU with more than 16GB memory may be required is using tensorflow-gpu. If this is not available, then the CPU-only version of tensorflow should also work, but it performance may be slow. To run the ventricle segmentation, the following command can be used:

scout segment ventricle syto_down6x.tif path/to/unet.h5 segment_ventricles.tif -t 0.5 -v

where path/to/unet.h5 is a path to the U-Net model weights and segment_ventricles.tif is the resulting binary ventricle segmentation image. The -t 0.5 argument specifies a probability threshold for obtaining a binary image from the U-Net model prediction.

Foreground segmentation

The overall organoid can be segmented by blurring the downsampled nuclear stain and thresholding the result, which can be done using the following command:

scout segment foreground syto_down6x.tif segment_foreground.tif -v -t 0.02 -g 8 4 4

where segment_foreground.tif is a newly created 3D binary TIFF image containing the overall organoid shape. The argument -g 8 4 4 specifies the amount of gaussian smoothing along each the z, y, and x dimensions, and the argument -t 0.02 specifies the level at which to threshold the gaussian smoothed image.