tif_paths = glob.glob(patch_folder + '/*.tif')
tif_paths = random.sample(tif_paths, 500)
# Load over images
for tif_path in tif_paths:
"""Run segmentation on an image."""
print(tif_path)
folder, tif_file = os.path.split(tif_path)
ndpi_id, x, y, w, h = tif_file[:-4].split('_') # Find patch location
x, y, w, h = int(x), int(y), int(w), int(h)
# Load NDPI slide
ndpi_path = os.path.join(DIR_cLEAN2, ndpi_id + '.ndpi')
ndpi_slide = NDPI_Slide(ndpi_path)
# Load image and run detection
image = ndpi_slide.read_region((x, y), 0, (patch_size, patch_size))
height, width, depth = image.shape
image = cv2.resize(image, (int(width / 2), int(height / 2)))
detection = model.detect([image])
if len(detection) == 0:
continue
results = detection[0]
# save segmentation for visualization
# vis_path = tif_path.replace('.tif', '.jpg')
# visualize.display_instances(
# image, results['rois'], results['masks'], results['class_ids'],
# 'nucleus', results['scores'],
# We process only slides, which has blue ink rois
roi_paths = glob.glob(DIR_roi_level0 + "/*.txt")
# roi_paths = glob.glob(DIR_roi_level0+"/12XS12147.txt")
for roi_path in roi_paths:
_, roi_file = os.path.split(roi_path)
ndpi_id = roi_file[:-4]
# read NDPA annotation
ndpa_annotation_path = os.path.join(DIR_annotated, ndpi_id + '.ndpi.ndpa')
if not os.path.exists(ndpa_annotation_path):
print('Can not find ' + ndpa_annotation_path)
ink_ndpi_slide_path = os.path.join(DIR_ink, ndpi_id + '.ndpi')
ink_slide = NDPI_Slide(ink_ndpi_slide_path)
polygons_ink, rects_ink, title_ink = ink_slide.read_ndpa_annotation(
ndpa_annotation_path)
# ink_patch = ink_slide.read_region((rects_ink[0][0], rects_ink[0][1]), 0, (rects_ink[0][2], rects_ink[0][3]))
# skimage.io.imsave('ink_patch.bmp', ink_patch)
# read ink2cLEARN align parameters
align_path = os.path.join(DIR_image_align_ink2clean, ndpi_id + '.txt')
align_para = np.loadtxt(align_path, delimiter=",")
# covert to cLEAN2 image coord
rects_cLEAN2 = [] # stores NDPA abnormal annotation in cLEARN2 coordinate
x_offset = align_para[0, 2] * 4 # align parameter is at Level 2 resolution
y_offset = align_para[
1,
2] * 4 # We work on Level 0 resolution, so there is a 2^2 difference in scale
def crop_patch():
ndpi_path = os.path.join(DIR_cLEAN2, '12XS00147.ndpi')
ndpi_slide = NDPI_Slide(ndpi_path)
roi_image = ndpi_slide.read_region((30720, 28672), 0, (2048, 2048))
skimage.io.imsave('patch.jpg', roi_image)
DIR_roi_level0, \
DIR_patch_level0
#
# Crop level 0 roi and save as TIF images
# Totally 6 slides have no blue ink ROI, we ignore them
#
from cercyt.shared.NDPI_Slide import NDPI_Slide
for coord_path in glob.glob(DIR_roi_level0 + '/*.txt'):
# Read coord info
coord_file = open(coord_path)
lines = coord_file.readlines()
coord_file.close()
# Read NDPI slide
folder, file = os.path.split(coord_path)
ndpi_id = file[:-4]
ndpi_path = os.path.join(DIR_cLEAN2, ndpi_id + '.ndpi')
ndpi_slide = NDPI_Slide(ndpi_path)
# Crop the ROIs
for line in lines:
words = line.strip().split(',')
x, y, w, h = int(words[1]), int(words[2]), int(words[3]), int(words[4])
roi_image = ndpi_slide.read_region((x, y), 0, (w, h))
roi_file = '{0}_{1}_{2}_{3}_{4}.tif'.format(ndpi_id, x, y, w, h)
roi_path = os.path.join(DIR_patch_level0, roi_file)
skimage.io.imsave(roi_path, roi_image)
model = modellib.MaskRCNN(mode="inference", config=config, model_dir='')
model.load_weights(model_path, by_name=True)
# Load over images
tif_paths = glob.glob(patch_folder + '/*.tif')
for tif_path in tif_paths:
"""Run segmentation on an image."""
print(tif_path)
folder, tif_file = os.path.split(tif_path)
ndpi_id, x, y, w, h = tif_file[:-4].split('_') # Find patch location
x, y, w, h = int(x), int(y), int(w), int(h)
# Load NDPI slide
ndpi_path = os.path.join(DIR_cLEAN2, ndpi_id + '.ndpi')
ndpi_slide = NDPI_Slide(ndpi_path)
# Load NDPA annotation
ndpa_annotation_path = os.path.join(DIR_annotated, ndpi_id + '.ndpi.ndpa')
polygons, rects, titles = ndpi_slide.read_ndpa_annotation(
ndpa_annotation_path)
# Pick the one annotation corresponding to this patch
for i, rect in enumerate(rects):
if len(intersection(rect, (x, y, w, h))) is not 0:
polygon = polygons[i]
break
# the top left corner of the patch to run mrcnn, we work at level 1 resolution
x_0, y_0 = float(x) + float(w) / 2.0, float(y) + float(h) / 2.0
location = int(x_0 - patch_size / 2 + 0.5), int(y_0 - patch_size / 2 + 0.5)
DataInfo, \
FILE_DATA_INFO, DIR_annotated, DIR_ink, DIR_image_align_ink2clean, DIR_cLEAN2, DIR_G_Tom_Patch_normal, \
DIR_G_Tom_Patch_malignancy, DIR_G_Tom_Patch_abnormal
from cercyt.shared.NDPI_Slide import NDPI_Slide
data_info = DataInfo(FILE_DATA_INFO)
slide_ids = data_info.get_abnormal_slide_ids()
for ndpi_id in slide_ids:
# read NDPA annotation
ndpa_annotation_path = os.path.join(DIR_annotated, ndpi_id + '.ndpi.ndpa')
if not os.path.exists(ndpa_annotation_path):
print('Can not find ' + ndpa_annotation_path)
ink_ndpi_slide_path = os.path.join(DIR_ink, ndpi_id + '.ndpi')
ink_slide = NDPI_Slide(ink_ndpi_slide_path)
polygons_ink, rects_ink, title_ink = ink_slide.read_ndpa_annotation(
ndpa_annotation_path)
# ink_patch = ink_slide.read_region((rects_ink[0][0], rects_ink[0][1]), 0, (rects_ink[0][2], rects_ink[0][3]))
# skimage.io.imsave('ink_patch.bmp', ink_patch)
# read ink2cLEARN align parameters
align_path = os.path.join(DIR_image_align_ink2clean, ndpi_id + '.txt')
align_para = np.loadtxt(align_path, delimiter=",")
# covert to cLEAN2 image coord
rects_cLEAN2 = [] # stores NDPA abnormal annotation in cLEARN2 coordinate
x_offset = align_para[0, 2] * 4 # align parameter is at Level 2 resolution
y_offset = align_para[
1,
2] * 4 # We work on Level 0 resolution, so there is a 2^2 difference in scale
# Collect some patches from normal slides
slides = data_info.get_normal_cell_patch_classification_train_slide_ids() + \
data_info.get_normal_cell_patch_classification_test_slide_ids()
dir_to_save = os.path.join(DIR_nuclei_segmentation, 'normal')
for slide_id in slides:
patch_folder = os.path.join(DIR_tile_patches, slide_id)
tif_paths = glob.glob(patch_folder + '/*.tif')
tif_paths = random.sample(tif_paths, 20)
# Load over images
for tif_path in tif_paths:
print(tif_path)
folder, tif_file = os.path.split(tif_path)
ndpi_id, x, y, w, h = tif_file[:-4].split('_') # Find patch location
x, y, w, h = int(x), int(y), int(w), int(h)
# Load NDPI slide
ndpi_path = os.path.join(DIR_cLEAN2, ndpi_id + '.ndpi')
ndpi_slide = NDPI_Slide(ndpi_path)
# Load image and save results
image = ndpi_slide.read_region((x, y), 0, (patch_size, patch_size))
cell_path = os.path.join(
dir_to_save,
'{0}_{1}_{2}_{3}_{4}.bmp'.format(ndpi_id, x, y, patch_size,
patch_size))
skimage.io.imsave(cell_path, image)
config = NucleusInferenceConfig()
config.display()
model = modellib.MaskRCNN(mode="inference", config=config, model_dir='')
model.load_weights(model_path, by_name=True)
image_folder = os.path.join(DIR_nuclei_segmentation, 'normal')
image_paths = glob.glob(image_folder + '/*.bmp')
for image_path in image_paths:
folder, tif_file = os.path.split(image_path)
ndpi_id, x, y, w, h = tif_file[:-4].split('_') # Find patch location
x, y, w, h = int(x), int(y), int(w), int(h)
# Load NDPI slide
ndpi_path = os.path.join(DIR_cLEAN2, ndpi_id + '.ndpi')
ndpi_slide = NDPI_Slide(ndpi_path)
# Load image and run detection
image = ndpi_slide.read_region((x, y), 0, (h, w))
height, width, depth = image.shape
image = cv2.resize(image, (int(width / 2), int(height / 2)))
detection = model.detect([image])
if len(detection) == 0:
continue
results = detection[0]
# save segmentation for visualization
vis_path = image_path.replace('.bmp', '.jpg')
visualize.display_instances(image,
results['rois'],