Question about Tuning Parameters to get Voronoi Plots

[DattaSujay]

I had a query regarding tuning the model for the TissueNet dataset taken from https://datasets.deepcell.org/data. The current shape of training and testing data is (10404,256,256,2)

The model was run as follows :

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

from deepcell.model_zoo.panopticnet import PanopticNet

classes = {
    'inner_distance': 1,  # inner distance
    'outer_distance': 1,  # outer distance
    'fgbg': 2,  # foreground/background separation
}

model = PanopticNet(
    backbone='resnet50',
    input_shape=X_train.shape[1:],
    norm_method='std',
    num_semantic_classes=classes)

from tensorflow.keras.optimizers import SGD, Adam
from deepcell.utils.train_utils import rate_scheduler

model_name = 'watershed_centroid_nuclear_general_std'

n_epoch = 5  # Number of training epochs
test_size = .20  # % of data saved as test
norm_method = 'whole_image'  # data normalization

lr = 1e-5
optimizer = Adam(learning_rate=lr, clipnorm=0.001)
lr_sched = rate_scheduler(lr=lr, decay=0.99)

batch_size = 1

min_objects = 3  # throw out images with fewer than this many objects
from deepcell import image_generators
from deepcell.utils import train_utils

transforms = list(classes.keys())
transforms_kwargs = {'outer-distance': {'erosion_width': 0}}

# use augmentation for training but not validation
datagen = image_generators.SemanticDataGenerator(
    rotation_range=180,
    shear_range=0,
    zoom_range=(0.75, 1.25),
    horizontal_flip=True,
    vertical_flip=True)

datagen_val = image_generators.SemanticDataGenerator(
    rotation_range=0,
    shear_range=0,
    zoom_range=0,
    horizontal_flip=0,
    vertical_flip=0)

train_data = datagen.flow(
    {'X': X_train, 'y': y_train},
    seed=seed,
    transforms=transforms,
    transforms_kwargs=transforms_kwargs,
    min_objects=min_objects,
    batch_size=batch_size)

val_data = datagen_val.flow(
    {'X': X_test, 'y': y_test},
    seed=seed,
    transforms=transforms,
    transforms_kwargs=transforms_kwargs,
    min_objects=min_objects,
    batch_size=batch_size)

from matplotlib import pyplot as plt

inputs, outputs = train_data.next()

img = inputs[0]
inner_distance = outputs[0]
outer_distance = outputs[1]
fgbg = outputs[2]

fig, axes = plt.subplots(1, 4, figsize=(15, 15))

axes[0].imshow(img[..., 0])
axes[0].set_title('Source Image')

axes[1].imshow(inner_distance[0, ..., 0])
axes[1].set_title('Inner Distance')

axes[2].imshow(outer_distance[0, ..., 0])
axes[2].set_title('Outer Distance')

axes[3].imshow(fgbg[0, ..., 0])
axes[3].set_title('Foreground/Background')

plt.show()

from timeit import default_timer

start = default_timer()
test_images = prediction_model.predict(X_test)
watershed_time = default_timer() - start

print('Watershed segmentation of shape', test_images[0].shape, 'in', watershed_time, 'seconds.')

import time

from matplotlib import pyplot as plt
import numpy as np

from skimage.feature import peak_local_max

from deepcell_toolbox.deep_watershed import deep_watershed

index = np.random.choice(X_test.shape[0])
print(index)

fig, axes = plt.subplots(1, 4, figsize=(20, 20))

masks = deep_watershed(
    test_images,
    min_distance=10,
    detection_threshold=0.1,
    distance_threshold=0.01,
    exclude_border=False,
    small_objects_threshold=0)

# calculated in the postprocessing above, but useful for visualizing
inner_distance = test_images[0]
outer_distance = test_images[1]

coords = peak_local_max(
    inner_distance[index],
    min_distance=10,
    threshold_abs=0.1,
    exclude_border=False)

# raw image with centroid
axes[0].imshow(X_test[index, ..., 0])
axes[0].scatter(coords[..., 1], coords[..., 0],
                color='r', marker='.', s=10)

axes[1].imshow(inner_distance[index, ..., 0], cmap='jet')
axes[2].imshow(outer_distance[index, ..., 0], cmap='jet')
axes[3].imshow(masks[index, ...], cmap='jet')

plt.show()

How to tune the parameters in a way that we obtain plots where the red dots align perfectly on the segmented cell data from which Voronoi plots can be generated? Will reshaping the data solve the issue?

[rossbar]

It is difficult to say - there is a lot going on with the example. It sounds like an image alignment issue. At first glance I doubt this is a matter of parameter tuning, but rather more likely related to what is being compared. I would recommend focusing on the shapes and dimensions of the inputs/outputs at each stage of the analysis, and careful review of the docstrings to ensure you understand any transforms that are applied to the images.