Finding Spots with LocalMaxPeakFinder

RNA FISH spots are usually bright point spread functions in a greyscale image. Rolonies, which are rolling-circle amplicons produced in certain assays (e.g. in situ sequencing), are approximately 1-um diameter Gaussian spots. Generally, the recommended FindSpotsAlgorithm to use in a starfish pipeline is BlobDetector because it accounts for the intensity profile of a spot rather than just thresholding pixel values. But for some images BlobDetector may not be satisfactory so starfish also provides alternatives.

LocalMaxPeakFinder finds spots by finding the optimal threshold to binarize image into foreground (spots) and background and then using skimage.feature.peak_local_max() to find local maxima or peaks in the foreground. If spots are being detected in the dark areas, increase the stringency value to increase the threshold. Using LocalMaxPeakFinder requires knowledge of expected foreground sizes since it uses min_obj_area and max_obj_area to filter connected foreground components before finding peaks. On the low end, min_obj_area should be set to the area of one spot so that smaller foreground objects are removed. On the high end, max_obj_area can be set to np.inf and decreased if necessary. min_distance is another threshold to filter out noise by limiting the distance between peaks that can be detected. The recommended way to set parameters is to take a representative image and visually assess results. Each peak is counted as a spot with radius equal to one pixel.


Running LocalMaxPeakFinder on 3-Dimensional images with is_volume=True can be extremely slow.


LocalMaxPeakFinder is not compatible with cropped data sets.

# Load osmFISH experiment
from starfish import FieldOfView, data
from starfish.image import Filter
from starfish.spots import DecodeSpots, FindSpots
from starfish.types import Axes, TraceBuildingStrategies
experiment = data.osmFISH(use_test_data=True)
imgs = experiment["fov_000"].get_image(FieldOfView.PRIMARY_IMAGES)

# filter raw data
filter_ghp = Filter.GaussianHighPass(sigma=(1,8,8), is_volume=True)
filter_laplace = Filter.Laplace(sigma=(0.2, 0.5, 0.5), is_volume=True), in_place=True), in_place=True)

# z project
max_imgs = imgs.reduce({Axes.ZPLANE}, func="max")

# run LocalMaxPeakFinder on max projected image
lmp = FindSpots.LocalMaxPeakFinder(
spots =

Gallery generated by Sphinx-Gallery