def test_iss_pipeline():
np.random.seed(2)
synthesizer = SyntheticData(n_spots=5)
codebook = synthesizer.codebook()
true_intensities = synthesizer.intensities(codebook=codebook)
image = synthesizer.spots(intensities=true_intensities)
dots_data = image.max_proj(Indices.HYB, Indices.CH, Indices.Z)
dots = ImageStack.from_numpy_array(
dots_data.reshape((1, 1, 1, *dots_data.shape)))
wth = WhiteTophat(disk_size=15)
wth.filter(image)
wth.filter(dots)
fsr = FourierShiftRegistration(upsampling=1000, reference_stack=dots)
fsr.register(image)
min_sigma = 1.5
max_sigma = 5
num_sigma = 10
threshold = 1e-4
gsd = GaussianSpotDetector(
min_sigma=min_sigma,
max_sigma=max_sigma,
num_sigma=num_sigma,
threshold=threshold,
blobs_stack=dots,
measurement_type='max',
)
intensities = gsd.find(hybridization_image=image)
assert intensities.shape[0] == 5
codebook.decode_euclidean(intensities)
def test_medium_synthetic_stack():
np.random.seed(0)
n_z = 40
height = 300
width = 400
sigma = 2
sd = SyntheticData(
n_round=4,
n_ch=4,
n_z=n_z,
height=height,
width=width,
n_spots=100,
n_codes=10,
point_spread_function=(sigma, sigma, sigma),
)
codebook = sd.codebook()
intensities = sd.intensities(codebook=codebook)
# some spots won't be detected properly because they will spread outside the image when blurred,
# so we'll remove those from intensities before we generate spots.
spot_radius = sigma * np.sqrt(
2) # this is the radius of the spot in pixels
valid_z = np.logical_and(intensities.z.values > spot_radius,
intensities.z.values < (n_z - spot_radius))
valid_y = np.logical_and(intensities.y.values > spot_radius,
intensities.y.values < (height - spot_radius))
valid_x = np.logical_and(intensities.x.values > spot_radius,
intensities.x.values < (width - spot_radius))
valid_locations = valid_z & valid_y & valid_x
intensities = intensities[np.where(valid_locations)]
spots = sd.spots(intensities=intensities)
blobs_image = spots.max_proj(Indices.CH, Indices.ROUND)
gsd = GaussianSpotDetector(min_sigma=1,
max_sigma=4,
num_sigma=5,
threshold=1e-4)
calculated_intensities = gsd.run(spots, blobs_image=blobs_image)
calculated_intensities = codebook.metric_decode(calculated_intensities,
max_distance=1,
min_intensity=0,
norm_order=2)
# spots are detected in a different order that they're generated; sorting makes comparison easy
sorted_intensities = intensities.sortby(
[Indices.Z.value, Indices.Y.value, Indices.X.value])
sorted_calculated_intensities = calculated_intensities.sortby(
[Indices.Z.value, Indices.Y.value, Indices.X.value])
# verify that the spots are all detected, and decode to the correct targets
assert np.array_equal(
sorted_intensities[Features.TARGET].values,
sorted_calculated_intensities[Features.TARGET].values)
def synthetic_dataset_with_truth_values():
from starfish.util.synthesize import SyntheticData
np.random.seed(2)
synthesizer = SyntheticData(n_spots=5)
codebook = synthesizer.codebook()
true_intensities = synthesizer.intensities(codebook=codebook)
image = synthesizer.spots(intensities=true_intensities)
return codebook, true_intensities, image
def test_medium_synthetic_stack():
np.random.seed(0)
n_z = 40
height = 300
width = 400
sigma = 2
sd = SyntheticData(
n_hyb=4,
n_ch=4,
n_z=n_z,
height=height,
width=width,
n_spots=100,
n_codes=10,
point_spread_function=(sigma, sigma, sigma),
)
codebook = sd.codebook()
intensities = sd.intensities(codebook=codebook)
# some spots won't be detected properly because they will spread outside the image when blurred,
# so we'll remove those from intensities before we generate spots.
spot_radius = sigma * np.sqrt(
2) # this is the radius of the spot in pixels
valid_z = np.logical_and(intensities.z.values > spot_radius,
intensities.z.values < (n_z - spot_radius))
valid_y = np.logical_and(intensities.y.values > spot_radius,
intensities.y.values < (height - spot_radius))
valid_x = np.logical_and(intensities.x.values > spot_radius,
intensities.x.values < (width - spot_radius))
valid_locations = valid_z & valid_y & valid_x
intensities = intensities[np.where(valid_locations)]
spots = sd.spots(intensities=intensities)
gsd = GaussianSpotDetector(min_sigma=1,
max_sigma=4,
num_sigma=5,
threshold=1e-4,
blobs_stack=spots)
calculated_intensities = gsd.find(spots)
codebook.decode_euclidean(calculated_intensities)
# spots are detected in a different order that they're generated; sorting makes comparison easy
sorted_intensities = intensities.sortby('features')
sorted_calculated_intensities = calculated_intensities.sortby('features')
# verify that the spots are all detected, and decode to the correct genes
assert np.array_equal(
sorted_intensities.features.gene_name.values,
sorted_calculated_intensities.features.gene_name.values)
def test_fov_order():
data = SyntheticData()
codebook = data.codebook()
stack1 = ImageStack.synthetic_stack()
stack2 = ImageStack.synthetic_stack()
fovs = [
FieldOfView("stack2", {"primary": stack2}),
FieldOfView("stack1", {"primary": stack1})
]
extras = {"synthetic": True}
experiment = Experiment(fovs, codebook, extras)
assert "stack1" == experiment.fov().name
assert ["stack1", "stack2"] == [x.name for x in experiment.fovs()]
def test_round_trip_synthetic_data():
np.random.seed(0)
sd = SyntheticData(
n_ch=2,
n_round=3,
n_spots=1,
n_codes=4,
n_photons_background=0,
background_electrons=0,
camera_detection_efficiency=1.0,
gray_level=1,
ad_conversion_bits=16,
point_spread_function=(2, 2, 2),
)
codebook = sd.codebook()
intensities = sd.intensities(codebook=codebook)
spots = sd.spots(intensities=intensities)
spots_mp = spots.max_proj(Axes.CH, Axes.ROUND)
spots_mp_numpy = spots_mp._squeezed_numpy(Axes.CH, Axes.ROUND)
gsd = BlobDetector(min_sigma=1, max_sigma=4, num_sigma=5, threshold=0)
calculated_intensities = gsd.run(spots, blobs_image=spots_mp_numpy)
decoded_intensities = codebook.metric_decode(
calculated_intensities,
max_distance=1,
min_intensity=0,
norm_order=2
)
# applying the gaussian blur to the intensities causes them to be reduced in magnitude, so
# they won't be the same size, but they should be in the same place, and decode the same
# way
spot1, ch1, round1 = np.where(intensities.values)
spot2, ch2, round2 = np.where(calculated_intensities.values)
assert np.array_equal(spot1, spot2)
assert np.array_equal(ch1, ch2)
assert np.array_equal(round1, round2)
assert np.array_equal(
intensities.coords[Features.TARGET],
decoded_intensities.coords[Features.TARGET]
)
def test_round_trip_synthetic_data():
np.random.seed(0)
sd = SyntheticData(
n_ch=2,
n_hyb=3,
n_spots=1,
n_codes=4,
n_photons_background=0,
background_electrons=0,
camera_detection_efficiency=1.0,
gray_level=1,
ad_conversion_bits=16,
point_spread_function=(2, 2, 2),
)
codebook = sd.codebook()
intensities = sd.intensities(codebook=codebook)
spots = sd.spots(intensities=intensities)
gsd = GaussianSpotDetector(min_sigma=1,
max_sigma=4,
num_sigma=5,
threshold=0,
blobs_stack=spots)
calculated_intensities = gsd.find(spots)
codebook.decode_euclidean(calculated_intensities)
# applying the gaussian blur to the intensities causes them to be reduced in magnitude, so
# they won't be the same size, but they should be in the same place, and decode the same
# way
spot1, ch1, hyb1 = np.where(intensities.values)
spot2, ch2, hyb2 = np.where(calculated_intensities.values)
assert np.array_equal(spot1, spot2)
assert np.array_equal(ch1, ch2)
assert np.array_equal(hyb1, hyb2)
assert np.array_equal(
intensities.coords[intensities.Constants.GENE],
calculated_intensities.coords[intensities.Constants.GENE])
