def plot_regions_on_all_images():
k_list = []
for j in range(nb_images):
all_mask = segmentation(image_lab_list[j],
2,
2,
2,
brown_lab=np.array([16.36, 110.89, 53.06]),
blue_lab=np.array([18.10, 50.58, -70.44]),
white_lab=np.array([53.15, -18.15, 1.20]))
results_p_dir_bis = os.path.join(results_dir, "image", f'{j}')
os.makedirs(results_p_dir_bis, exist_ok=True)
regions_positive = outline_regions(
image_original_list[j][10:image_original_list[j].shape[0] - 10,
10:image_original_list[j].shape[1] - 10],
all_mask[0, :, :])
io.imsave(fname=os.path.join(results_p_dir_bis,
f'regions positive.jpg'),
arr=regions_positive)
regions_negative = outline_regions(
image_original_list[j][10:image_original_list[j].shape[0] - 10,
10:image_original_list[j].shape[1] - 10],
all_mask[1, :, :])
io.imsave(fname=os.path.join(results_p_dir_bis,
f'regions negative.jpg'),
arr=regions_negative)
regions_background = outline_regions(
image_original_list[j][10:image_original_list[j].shape[0] - 10,
10:image_original_list[j].shape[1] - 10],
all_mask[2, :, :])
io.imsave(fname=os.path.join(results_p_dir_bis,
f'regions background.jpg'),
arr=regions_background)
k = validation(all_mask[0, :, :], all_mask[1, :, :], all_mask[2, :, :],
ref_paths[j][0], ref_paths[j][1])
print('k = ', k)
k_list.append(k)
k_history[j].append(k)
print('k mean = ', np.mean(k_list))
k_history[-1].append(np.mean(k_list))
return 1 / np.mean(k_list)
def label_patches(self, patches, max_patches=None):
if max_patches is not None:
patches = itertools.islice(patches, max_patches)
for patch in patches:
assert isinstance(patch, Patch)
os.makedirs(root_dir(FOLDER_TEMP_DATA), exist_ok=True)
path_patch = root_dir(FOLDER_TEMP_DATA, 'patch.png')
path_outln = root_dir(FOLDER_TEMP_DATA, 'outln.png')
images = {
path_patch: patch.cropped_image(),
path_outln: outline_regions(image=patch.cropped_image(),
region_labels=patch.cropped_mask())
}
[imsave(fname=rel_path, arr=resize(img, output_shape=(600, 600))) for rel_path, img in images.items()]
labels = LABELS_Ki67 + ['Not well defined']
while patch.expert_label not in labels:
# noinspection PyTypeChecker
gui = buttonbox("Ki67 - Manual labelling", choices=labels,
images=list(images.keys()), run=False)
gui.ui.set_pos("+0+0")
patch.expert_label = gui.run()
print(patch.expert_label)
self.labelled_patches.append(patch)
arr=image)
logging.info('CLAHE')
image = apply_on_normalized_luminance(
lambda img: equalize_adapthist(img, clip_limit=0.02),
image_rgb=image)
io.imsave(fname=os.path.join(results_p_dir, f'01 03 - CLAHE.jpg'),
arr=image)
logging.info('K-means clustering')
image_flat = rgb2lab(image).reshape((-1, 3))
clustering = KMeans(n_clusters=CLUSTERING_NUM_CENTROIDS, random_state=0).fit(image_flat)
io.imsave(fname=os.path.join(results_p_dir, f'02 K-means - labels.jpg'),
arr=colormap(clustering.labels_.reshape(image.shape[0:2])))
io.imsave(fname=os.path.join(results_p_dir, f'02 K-means - regions.jpg'),
arr=outline_regions(image=original_image, region_labels=clustering.labels_.reshape(image.shape[0:2])))
io.imsave(fname=os.path.join(results_p_dir, f'02 K-means - average_color.jpg'),
arr=average_color(image=original_image, region_labels=clustering.labels_.reshape(image.shape[0:2])))
logging.info('Class separation')
# Positive mask: cluster with maximum on channel a
idx_positive_cluster = np.argmax(clustering.cluster_centers_[:, 1])
positive_mask = np.equal(clustering.labels_, idx_positive_cluster).reshape(image.shape[0:2])
# Negative mask: cluster with minimum on channel b
idx_negative_cluster = np.argmin(clustering.cluster_centers_[:, 2])
negative_mask = np.equal(clustering.labels_, idx_negative_cluster).reshape(image.shape[0:2])
# Visualize
io.imsave(fname=os.path.join(results_p_dir, f'03 Positives.jpg'),
arr=outline_regions(image=original_image, region_labels=positive_mask))
io.imsave(fname=os.path.join(results_p_dir, f'03 Negatives.jpg'),
arr=outline_regions(image=original_image, region_labels=negative_mask))
logging.info('Visualizing classification')
classification_colored = visualize_classification(classification)
io.imsave(fname=os.path.join(results_p_dir, '04 04 Classification colored.jpg'),
arr=classification_colored)
logging.info('Creating some patches')
c1 = PatchClassifier()
patches = list(c1.patches(image=image, region_labels=segmentation_idcs))
logging.info('Labelling patches')
c1.label_patches(patches=patches, max_patches=MAX_PATCHES_PER_IMAGE)
logging.info('Saving patches')
c1.save_labelled_patches(training_label=execution_id)
logging.info('Retrieving patches')
c2 = PatchClassifier()
c2.load_labelled_patches(training_label=execution_id)
logging.info('Visualizing retrieved patches')
for idx_patch, patch in enumerate(c2.labelled_patches):
io.imsave(fname=os.path.join(results_p_dir, f'05 {idx_patch:02d} Patch image classification.jpg'),
arr=colormap(crop(image=segmentation_idcs, bounding_box=patch.bounding_box)))
io.imsave(fname=os.path.join(results_p_dir, f'05 {idx_patch:02d} Patch image outlined.jpg'),
arr=outline_regions(patch.cropped_image(), patch.cropped_mask()))
io.imsave(fname=os.path.join(results_p_dir, f'05 {idx_patch:02d} Patch image.jpg'),
arr=patch.cropped_image())
io.imsave(fname=os.path.join(results_p_dir, f'05 {idx_patch:02d} Patch mask.jpg'),
arr=colormap(patch.cropped_mask()))
logging.info(
'Positive separation (k-means clustering on `a channel`)')
channel_a = image_lab[:, :, 1]
clustering = KMeans(n_clusters=2,
random_state=0).fit(channel_a.reshape(-1, 1))
idx_positive_cluster = np.argmax(clustering.cluster_centers_)
positive_mask = np.equal(clustering.labels_,
idx_positive_cluster).reshape(
channel_a.shape[0:2])
io.imsave(fname=os.path.join(results_p_dir,
f'03 K-means - Positives - Mask.jpg'),
arr=colormap(positive_mask))
io.imsave(fname=os.path.join(
results_p_dir, f'03 K-means - Positives - regions.jpg'),
arr=outline_regions(image=image,
region_labels=positive_mask))
io.imsave(fname=os.path.join(
results_p_dir, f'03 K-means - Positives - average color.jpg'),
arr=average_color(image=image,
region_labels=positive_mask))
logging.info(
'Negative separation (k-means clustering on `b channel`)')
channel_b = image_lab[:, :, 2]
nonpositive_mask = np.logical_not(positive_mask)
clustering = KMeans(n_clusters=2, random_state=0).fit(
channel_b[nonpositive_mask].reshape(-1, 1))
idx_negative_cluster = np.argmin(clustering.cluster_centers_)
negative_mask_wrt_nonpositive = np.equal(clustering.labels_,
idx_negative_cluster)
negative_mask = np.full(shape=positive_mask.shape,
image_lab = rgb2lab(image)
logging.info('Resizing')
resize_factor = 8
image_lab = resize(image_lab, (int(image.shape[0] / resize_factor),
(image.shape[1] / resize_factor)),
anti_aliasing=True)
all_mask = segmentation(image_lab,
brown_lab=np.array([16.36, 110.89, 53.06]),
blue_lab=np.array([18.10, 50.58, -70.44]),
white_lab=np.array([53.15, -18.15, 1.20]),
sigma=0.454)
regions_positive = outline_regions(
resized_original[5:resized_original.shape[0] - 5,
5:resized_original.shape[1] - 5],
all_mask[0, :, :])
io.imsave(fname=os.path.join(results_p_dir,
f'regions positive.jpg'),
arr=regions_positive)
regions_negative = outline_regions(
resized_original[5:resized_original.shape[0] - 5,
5:resized_original.shape[1] - 5],
all_mask[1, :, :])
io.imsave(fname=os.path.join(results_p_dir,
f'regions negative.jpg'),
arr=regions_negative)
regions_background = outline_regions(
resized_original[5:resized_original.shape[0] - 5,
5:resized_original.shape[1] - 5],
all_mask[2, :, :])
n = randint(0, 4) # to try the new value on one of the five images
masks = segmentation(image_lab_list[n],
2,
2,
2,
brown_lab=dictionary_arguments['brown_lab'],
blue_lab=dictionary_arguments['blue_lab'],
white_lab=dictionary_arguments['white_lab'])
results_p_dir_bis = os.path.join(results_dir, "image", param_name,
f'{result.x}')
os.makedirs(results_p_dir_bis, exist_ok=True)
regions_positive = outline_regions(
image_original_list[n][10:image_original_list[n].shape[0] - 10,
10:image_original_list[n].shape[1] -
10], masks[0, :, :])
io.imsave(fname=os.path.join(results_p_dir_bis,
f'regions positive.jpg'),
arr=regions_positive)
regions_negative = outline_regions(
image_original_list[n][10:image_original_list[n].shape[0] - 10,
10:image_original_list[n].shape[1] -
10], masks[1, :, :])
io.imsave(fname=os.path.join(results_p_dir_bis,
f'regions negative.jpg'),
arr=regions_negative)
regions_background = outline_regions(
image_original_list[n][10:image_original_list[n].shape[0] - 10,
10:image_original_list[n].shape[1] -
10], masks[2, :, :])